FIDO Biometrics Requirements

3.4.2.1. False Reject Rate (FRR) - BCC

For BioLevel 1 and BioLevel 2, the False Reject Rate SHALL meet the requirement of less than 7:100 for the upper bound of a 80% confidence interval. For BioLevel 1 and BioLevel 2, self-attestation of at least 5% false reject rate is also required and described in Section Documented Self-Attestation FRR.

For BioLevel 1+ and BioLevel 2+, the False Reject Rate SHALL meet the requirement of less than 5:100 for the upper bound of an 80% confidence interval.

3.4.3.1. False Accept Rate (FAR) - BCC

For BioLevel 1 and BioLevel 1+, the False Accept Rate SHALL meet the requirement of less than 1:100 for the upper bound of an 80% confidence interval. For BioLevel 1 and BioLevel 1+, documented self-attestation of at least 1/10000 false accept rate is also required and described in Section Documented Self-Attestation FAR.

For BioLevel 2 and BioLevel 2+, the False Accept Rate SHALL meet the requirement of less than 1:10,000 for the upper bound of an 80% confidence interval.

3.4.4. MultiBiometric Performance Testing - BCC Only

For TOEs that collect all biometric data for fusion prior to a decision, Testing SHALL be performed as written where subjects SHALL provide all modalities as part of authentication. FAR and FRR for the combined decision remain the same.

For TOEs that operate based on sequential fusion (i.e., where a decision resulting from initial biometric(s) determine whether subsequent biometric(s) need to be collected), the laboratory SHALL create a test protocol. If it is up to the user which modality goes first, a test plan SHALL be designed based on the information the vendor provides regarding the TOE and approved by the FIDO Biometric Secretariat. For example, the Laboratory MAY randomly assign the order of the modalities for each transaction. If it is up to the TOE which modality goes first, a test plan SHALL be designed based on the information the vendor provides regarding the TOE and approved by the FIDO Biometric Secretariat.

Methods for computing FAR and FRR remain the same and are based on the combined final decision.

For TOE that utilize different fusion approaches depending upon the environment, the Laboratory SHALL test each of these algorithms similar to Section Test Environment.

Note: For example, if face modality is always captured first, followed by fingerprint, only if needed, the test plan will consider face first. In another example, if the TOE chooses which modality goes first, the Laboratory shall consider this in the test plan.

3.4.5.1. Documented Self-Attestation FAR - BCC

Dcoumented self-attestation for FAR is optional for BioLevel 1+ and BioLevel 2+, but required for BioLevel 1 and BioLevel 2.

3.4.6.1. Documented Self-Attestation FRR - BCC

Self-attestation for FRR is optional for BioLevel 2 and BioLevel 2+, but required for BioLevel 1 and BioLevel 1+.

3.4.7. Maximum Number of Biometric References from Multiple Fingers - BCC Only

Requirement

If a subject enrolls multiple fingers (e.g., index and thumb) and uses them interchangeably (i.e., one OR the other), the FAR increases where the FAR for two fingers enrolled is approximately twice the FAR for one finger enrolled. This section describes the process such that a biometric system can be certified to operate with two or more enrolled fingers. Other biometric modalities where this may apply are described in notes below.

If the analysis below is not performed, the maximum number of biometrics references SHALL default to one.

The vendor SHALL declare the maximum number of different fingers which can be enrolled. The FAR associated with multiple biometric references (FAR_MT) SHALL be calculated according to the following and SHALL not be greater than 1:10,000.

FAR_MT =1 - ( (1-FAR_SA)^B )

B=Max # Biometric References FAR_SA=Self-attested FAR verified by FIDO according to Section Self-Attestation FAR (Optional).

At the time of FIDO authenticator certification, the maximum number of biometric references which meet the FAR requirement MAY be stored in the meta-data and SHALL NOT be greater than the maximum number verified during biometric certification, according to the above. Self-attested FAR in the meta-data SHALL be based on the single biometric reference FAR.

Note: Some iris systems may enroll each eye separately and allow successful verification even if only one eye is presented. Eyes can be considered in place of fingers for this section, if applicable to the TOE.

Note: The same process may be used for other modalities which have a similar property, i.e., where multiple parts of the body can be used interchangeably, e.g., palm veins for right and left hand. The vendor SHALL submit how this property may apply the modality of the TOE. The FIDO laboratory SHALL use the same process to assess the maximum number of biometric references.

3.5.1.1. Impostor Attack Presentation Accept Rate (IAPAR) - BCC

A TOE with an IAPAR of less than or equal to 15% will meet BioLevel 1 and BioLevel 1+ requirements. A TOE with an IAPAR of less than or equal to 7% will meet BioLevel 2 and BioLevel 2+ requirements.

The threshold for the level and the level achieved is stored in the FIDO Metadata as well as the FIDO Biometric Component Certificate.

3.5.2.1. Rate Limits - BCC

Additional requirements in the FIDO Authenticator Security Requirements may impact the biometric TOE under evaluation herein. Those are tested as part of FIDO Authenticator Certification. As part of these requirements, FIDO Authenticators are required to rate-limit user verification attempts per FIDO Authenticator Security Requirements, Requirement 3.9.

For the purposes of biometric certification testing, rate limiting SHOULD be turned off.

For BCC program, this optional certification SHALL only be available for TOE that are undergoing certification at Levels 1+ and 2+ with at least 245 subjects.

4.1. Common Test Harness - BCC

For each operating point to be evaluated, the Vendor SHALL provide a biometric system component of the FIDO authenticator which has, at a minimum, the items described in the next sections.

1. Configurable Enrollment system which:
   1. Selects the operating point(s) to be evaluated.
   2. Has enrollment hardware/software as will be executed by the FIDO authenticator.
   3. Includes a biometric data capture sensor and enrollment software.
   4. Can clear an enrollment.
   5. Can store an enrollment from acquired biometric sample(s) for use in an on-line verification evaluation.
   6. Can provide enrollment biometric references from acquired biometric sample(s) defined as “user’s store reference measure based on features extracted from enrollment samples” for use in off-line verification evaluation.
   7. Indicates a failure to enroll ([ISOIEC-19795-1], 4.6.1)
2. Configurable Biometric Verification on-line system which:
   1. Selects the operating point(s) to be evaluated.
   2. Has verification hardware/software to be executed by the FIDO authenticator.
   3. Includes a biometric data capture sensor, a biometric matcher, and a decision module.
   4. Captures features from an acquired biometric sample to be compared against a biometric reference.
   5. Makes an accept/reject decision at a specific operating point.
   6. Indicates an on-line failure to acquire ([ISOIEC-19795-1], 4.6.2).
   7. Indicates an on-line decision(accept or reject).
   8. Provides the decisive sample(s) of an online verification transaction, i.e., all data used to make the verification transaction decision (called a stored verification transaction). This will be used for off-line verification.
3. Configurable Biometric Verification off-line software, which:
   1. Selects the operating point(s) to be evaluated.
   2. Has verification software to be executed by the FIDO authenticator.
   3. Accepts a biometric reference and the stored verification transaction then performs matching in off-line batch mode.
   4. Provides a decision (accept or reject).
4. Logging capabilities, which:
   1. Records every interaction with the TOE.
   2. Allows the tester to manually add interactions (e.g., the fact that a tester just cleaned the sensor device)

   Note: For Enrollment, some vendors MAY use multiple samples per test subject (e.g., multiple impressions for a single finger). A biometric reference can be based on multiple stored samples. This SHOULD be opaque to the tester.

   Note: For a Stored Verification Transaction, the Test Harness SHALL store all attempts in a transaction. These stored attempts will be used for off-line verification testing.

5.1.1.1. Number of Subjects - BCC

For BioLevel 1 and BioLevel 2 certification, the minimum number of subjects for a test SHALL be 25, based on [ISOIEC-19795-1] and associated analysis in the Statistics and Test Size section of this document.

For BioLevel 1+ and BioLevel 2+ certification, the minimum number of subjects for a test SHALL be 245, based on [ISOIEC-19795-1] and associated analysis in the Statistics and Test Size section of this document.

For fingerprint, up to four different fingers from a single person can be considered as different test subjects. For the fingerprint biometrics, these SHALL be constrained to the index, thumb, or middle fingers, and SHALL be the same as was used for enrollment. A minimum of 123 unique persons SHALL be in the test crew.

Note: Having 123 test subjects will only require the use of 2 fingerprints per test subject at minimum. Allowing 4 fingers per test subject should allow the laboratory to acquire additional data if needed. It also allows to better align the test results of the Laboratory with the test results of a potential self attestation.

For eye-based biometrics, both the left and right eye can be considered as two different test subjects. A minimum of 123 unique persons SHALL be in the test crew.

Note: Two eyes cannot be considered as different test subjects if both eyes are enrolled at one time.

In the event there is an enrollment failure according to Enrollment Transaction Failures, an additional Subject SHALL be enrolled for each enrollment failure.

5.1.3.3.1. Rule of 3: FAR - BCC

For BioLevel 1 and BioLevel 2, we have 25 subjects, 300 combinations, and 1.61/300 leads to .536% FAR (<1%). For BioLevel 1+ and BioLevel 2+, we have 245 subjects, 29890 combinations, and 1.61/29890 leads to 0.00535% FAR (<0.01%).

5.1.3.4. Rule of 3: FRR

In the event that there are zero errors in the set of geniune comparisons, the TOE meets the FRR requirement on the basis of the Rule of 3.

Note: The "Rule of 3" method can be utilized to establish an upper bound if there are zero errors in the test, according to [ISOIEC-19795-1], Appendix B.1.1.

5.1.4. Test Visits

As this test is focused on False Accept Rate, collection from test subjects MAY occur in one visit.

5.1.5. Test Environment

In this context, it should be noted that the definition of the testing environment of the TOE, (which is based on the environment of the TOE described in the Allowed Integration Document), plays an important role in the context of the certification. For this reason, potential environment(s) shall also be described in the Allowed Integration Document ([BiometricAIDTempate]). Every certificate will identify the testing environment under which the biometric component has been tested.

The definition of the environment may also have an impact on the testing activities. Testing shall always be carried out under consideration of the intended environment. The test requirements in this document allow for a certain variation in environmental conditions. However, such variations have their limits. This could lead into a situation where the Laboratory SHALL perform a test multiple times or with a larger amount of test subjects if a TOE has a very diverse definition of environments.

The question of whether a specific intended environment will lead to additional requirements for testing has to be seen in the context of a specific Target of Evaluation and shall be discussed with the FIDO biometric secretariat during the review of the test plan. Environments that may lead to increased FRR (i.e., more inconvenience for the user) will not necessarily be evaluated as part of the testing plan. However, a testing plan may include multiple environments for cases where the TOE may have multiple configurations to address multiple environments.

For example, this could include (1) a different operating point (e.g., threshold for the matcher) for a noisy environment or (2) an NIR-only face recognition in low/no light (and visible light is used in normal light). If a TOE has multiple configurations that address different environments, then the TOE SHALL be tested for each configuration and the test plan SHALL incorporate the variations for the different environments that result in a different configuration.

5.1.7. Enrollment

Enrollment procedures SHALL be provided in writing to the Laboratory by the Vendor. These procedures SHALL be followed by the Test Crew. Instructions MAY be provided in any form, including interactive on screen guidance to the Test Subject. The Administrator SHALL record any FTE, if appropriate, along with any divergence from enrollment instructions that MAY have caused the failure.

5.2. Test Methods

Testing WILL be performed through a combination of Online and Offline Testing ([ISOIEC-19795-1]).

5.2.1. Pre-Testing Activities

Pre-test activities SHALL be performed according to [ISOIEC-19795-2]:

• Section 6.1.8 Pre-test procedures

• Section 6.1.8.1 Installation and validation of correct operation

5.2.2. Online Testing

This section will focus on Online Testing.

To facilitate estimation of a false accept rate, all biometric references and all stored verification transactions are stored to allow for offline computation of the FAR.

5.2.2.1. Online: Enrollment

Enrollment SHALL be performed according to [ISOIEC-19795-1], 7.3.

5.2.2.1.1. Pre-Enrollment

Before enrollment, test subjects MAY perform practice transactions.

5.2.2.1.2. Enrollment Transactions

Enrollment transactions SHALL be conducted without test operator guidance with the exception that the test operator may instruct the user to perform 5 mated transactions. Additionally, the operator is allowed to provide guidance as far as it concerns the test situation. Any kind of guidance SHALL be provided by the biometric authentication system/capture sensor in a way similar to the final application.

The enrollment process will be different depending on the biometric authentication system. This process MAY allow enrollment after one attempt, or MAY require multiple presentations and attempts. For testing, this process SHALL be similar to the final application.

5.2.2.1.3. Enrollment Transaction Failures

A failure to enroll SHALL be declared when the biometric authentication system is not able to generate a biometric reference for the test subjects after executing three enrollment transactions.

5.2.2.2. Online: Mated Verification Transaction

Mated verification transactions SHALL be performed according to [ISOIEC-19795-1], 7.4. This means that the following requirements SHALL be met:

Mated transaction data shall be collected in an environment, including noise, that closely approximates the target application. This test environment shall be consistent throughout the collection process. The motivation of test subjects, and their level of training and familiarity with the system, should also mirror that of the target application.

The collection process should ensure that presentation and channel effects are either uniform across all users, or randomly varying across users. If the effects are held uniform across users, then the same presentation and channel controls in place during enrollment should be in place for the collection of the test data. Systematic variation of presentation and channel effects between enrollment and test data will lead to results distorted by these factors. If the presentation and channel effects are allowed to vary randomly across test subjects, there shall be no correlation in these effects between enrollment and test sessions across all users.

In the ideal case between enrollment and the collection of test data, test subjects should use the system with the same frequency as the target application. However, this may not be a cost-effective use of the test crew. It may be better to forego any interim use, but allow re-familiarization attempts immediately prior to test data collection.

For systems that may adapt the biometric reference after successful verification, some interim use between enrolment and collection of mated attempt and transaction data may be appropriate. The amount of such use should be determined prior to data collection, and should be reported with results.

The sampling plan shall ensure that the data collected are not dominated by a small group of excessively frequent, but unrepresentative users.

Great care shall be taken to prevent data entry errors and to document any unusual circumstances surrounding the collection. Keystroke entry on the part of both test subjects and test administrators should be minimized. Data could be corrupted by impostors or mated users who intentionally misuse the system. Every effort SHALL be made by test personnel to discourage these activities; however, data SHALL NOT be removed from the corpus unless external validation of the misuse of the system is available.

Users are sometimes unable to give a usable sample to the system as determined by either the test administrator or the quality control module. Test personnel should record information on failure-to-acquire attempts where these would otherwise not be logged. The failure-to-acquire rate measures the proportion of such attempts, and is quality threshold dependent. As with enrollment, quality thresholds SHoULD be set in accordance with vendor advice.

Test data SHALL be added to the corpus regardless of whether or not it matches a biometric reference. Some vendor software does not record a measure from an enrolled user unless it matches the biometric reference. Data collection under such conditions would be severely biased in the direction of underestimating false non-match error rates. If this is the case, non-match errors shall be recorded by hand. Data shall be excluded only for predetermined causes independent of comparison scores.

All attempts, including failures-to-acquire, shall be recorded. In addition to recording the raw image data if practical, details SHALL be kept of the quality measures for each sample, if available and, in the case of online testing, the matching score or scores.

Note: Details for FIDO as they relate to the ISO requirements are discussed in the following sections.

5.2.2.2.1. Pre-Verification

Before mated verification transactions, test subjects MAY perform practice transactions.

5.2.2.2.2. Mated Verification Transaction

Test Subjects SHALL conduct 10 mated verification transactions. Mated verification transactions SHALL be conducted such that there is variability between transactions by introducing environmental or behavioral factors. Examples for each modality are provided below:

• Face: lighting, pose, expression<\li>
• Fingerprint: pressure, moisture, angle<\li>
• Iris: gaze, lighting, closed eyes<\li>

The chosen variabilities SHALL be documented as part of the FIDO Test plan and approved by the FIDO Biometric Secretariat. The FIDO Biometrics Secretariat shall ensure equivalency across laboratories in the chosen variabilities.

Mated verification transactions SHALL be conducted without test operator guidance. Any kind of guidance SHALL be provided by the biometric authentication system / capture sensor in a similar manner to the final application.

The verification process MAY be different depending on the biometric authentication system. This process MAY require multiple presentations. The verification transaction (maximum number of verification attempts and timeout period) for a TOE is defined in the Allowed Integration Document as described in Verification Transactions. A transaction SHOULD NOT exceed 30 seconds.

The authenticator vendor SHALL describe to the Accredited Biometric Laboratory what constitutes the start and the end of a verification transaction.

The test harness SHALL provide the decisive sample(s) of a transaction for off-line testing, i.e., all data used to make the verification transaction decision.

5.2.2.2.3. Mated Verification Errors

A failure to acquire SHALL be declared when the biometric authentication system is not able to capture and/or generate biometric features during a verification attempt (an FTA MAY happen per attempt). The on-line verification test harness SHALL indicate to the Laboratory when a failure to acquire has occurred.

Note: A failure to acquire will not be considered during off-line FAR testing.

A false rejection error SHALL be declared when the biometric authentication fails to authenticate the test subjects after executing the complete verification transaction.

The manner in which the Laboratory records failure to acquire, false rejects, and true accepts are left to the Laboratory, but SHALL be done automatically to avoid introducing human error.

5.2.2.2.4. FRR

False reject rate SHALL be calculated according to requirements in FRR and statistical analysis in Statistics and Test Size.

5.2.3. Offline Testing

Offline testing measures FAR and leverages all possible combinations between test subjects.

5.2.3.1. Offline: Software Validation

As the evaluation procedure might utilize online testing for the evaluation of a false reject rate and offline testing for the evaluation of false accept rates, it is important for the evaluation Laboratory to assure that the offline biometric functionality is functionally equivalent to the online biometric functionality. The evaluation Laboratory SHALL perform a series of mated verification transaction tests online and offline and make sure that the results are the same.

The on-line verification testing SHALL result in a sequence of stored verification transactions and decisions for every transaction that did not have a failure to acquire. The off-line verification testing SHALL run these stored verification transactions in the same order and SHALL result in the exact same sequence of decisions. The complete sequences SHALL be compared by the Laboratory to ensure their identity.

5.2.3.2. Offline: Non-mated Verification Transactions

5.2.3.2.1. Pre-Verification

To facilitate estimation of false accept rate, all enrollment transactions and verification transactions are stored to allow for offline computation of the FAR.

5.2.3.2.2. Non-mated Verification Transaction

The verification offline module provided by the vendor is used to compute all impostor (between person) combinations for estimating FAR.

Non-mated Verification transactions SHALL be performed according to [ISOIEC-19795-1], 7.6.1.1b, 7.6.1.2b, 7.6.1.3, 7.6.1.4, and 7.6.3.1. Different fingers or irises from the same person SHALL NOT be compared according to [ISOIEC-19795-1] 7.6.1.3.

5.2.3.2.3. Non-mated Verification Transaction Failures

The impostor verification process compares a biometric reference and a stored verification transaction from different persons.

For fingerprint, up to four different fingers from a single person can be considered as different test subjects. For eye-based biometrics, both the left and right eye can be considered as two different test subjects. However, impostor scores between two fingerprints or two irises from a single person SHALL be excluded from computation of the FAR.

A false accept error SHALL be declared if the stored verification transaction results in a match decision.

Note: It is not possible to obtain an FTA rate for FAR Offline Testing. FTAs are not considered in Offline Testing.

5.2.3.2.4. FAR

False accept rate SHALL be calculated according to requirements in FAR and statistical analysis in Statistics and Test Size.

6. Test Procedures for Presentation Attack Detection (PAD)

This section provides the testing plan for Presentation Attack (Spoof) Detection. It focuses on presentation attacks which require minimal expertise. The testing SHALL be performed by the FIDO-accredited independent testing Laboratory on the TOE provided by the vendor. The evaluation measures the Impostor Attack Presentation Accept Rate (IAPAR), as defined in ISO 30107 Part 3.

PAD Testing shall be completed by using the following approach.

6.1. Test Crew

The Test Crew are the Test Subjects gathered for evaluation.

6.1.1. Number of Subjects

Number of subjects for a test SHALL be 15.

For fingerprints, PAD testing SHALL be constrained to the index, thumb, or middle fingers of the test subject.

The same test subjects as used for FRR testing may be used for PAD testing.

In the event there is an enrollment failure according to Enrollment Transaction Failures, an additional Subject SHALL be enrolled for each enrollment failure.

6.1.2. Population

The population SHALL be representative of the target market in relationship to age and gender. Age and gender recommendations are taken from [ISOIEC-19795-5] for access control applications (Section 5.5.1.2 and 5.5.1.3). The following targets SHALL be used for age and gender. Minor deviations from these numbers may be acceptable if agreed by the FIDO biometric secretariat.

6.1.2.1. Age

Age Distribution Requirements

Age	Distribution
<18	0%
18-30	25-40%
31-50	25-40%
51+	25-40%

6.1.2.2. Gender

Gender Distribution Requirements

Gender	Distribution
Male	40-60%
Female	40-60%

6.1.2.3. Skin Tone

Groups based on skin tone SHALL be defined based on the Monk Scale [MonkSkinTone] and reported.

Note: As indicated in [ISOIEC-19795-1], ideally, the test subjects SHOULD be chosen at random from a population that is representative of the people who will use the system in the real application environment. In some cases, however, the test subjects do not accurately represent the real-world users. If the test crew comes from the vendor’s employee population, they MAY differ significantly from the target users in terms of educational level, cultural background, and other factors that can influence the performance with the chosen biometric system.

6.1.3. Test Visits

Collection from test subjects MAY occur in one visit.

6.1.4. Enrollment

Enrollment procedures SHALL be provided in writing to the Laboratory by the Vendor. These procedures SHALL be followed by the Test Crew. Instructions MAY be provided in any form, including interactive on screen guidance to the Test Subject. The Administrator SHALL record any FTE, if appropriate, with any divergence from enrollment instructions that MAY have caused the failure.

6.2. Test Methods

Testing will be performed through Online Testing using the Common Test Harness defined in Common Test Harness (Optional).

6.2.1. Pre-Testing Activities

Pre-test activities SHALL be performed according to [ISOIEC-19795-2]:

• Section 6.1.8 Pre-test procedures

• Section 6.1.8.1 Installation and validation of correct operation

6.2.2. Testing for PAD

This section will focus on PAD Testing.

6.2.3. Enrollment

Each subject SHALL be enrolled. Enrollment SHALL be performed according to ISO/IEC 19795-1, 7.3. Presentation attacks will be performed against this enrollment. Similar to FRR/FAR testing, enrollment transactions will be performed without operator guidance, and is flexible with regard to the vendor, in that it may allow multiple presentations for the enrollment.

The test laboratory SHALL also collect biometric characteristic data required for creating a Presentation Attack Instrument. For example, for fingerprint, a copy of the enrolled person’s fingerprint is needed and can be acquired via collection of a fingerprint image on a second fingerprint scanner or by leaving a latent print. The method used to acquire the biometric characteristic SHALL be consistent with the recipe of each Presentation Attack Instrument Species to be tested.

6.2.4. PAI Species

A Presentation Attack Instrument (PAI) is the device used when mounting a presentation attack. A PAI species is a set of PAIs which use the same production method, but only differ in the underlying biometric characteristic. Table 1 is a high level description of presentation attacks by level. The next section provides additional detail of PAI species for each modality.

The laboratory SHALL select PAI species appropriate for biometric modality of the TOE.

PAI Species are described at a high level for fingerprint, face, iris/eye, and voice. If the TOE is a different biometric modality than these, the vendor SHALL propose a set of PAI species for Levels A and B for FIDO’s approval. Upon FIDO approval, the test Laboratory can proceed with PAD evaluation.

6.2.5. PAD Evaluation with Presentation Attack Instruments (PAI)

Based upon the prior section, the test Laboratory SHALL select six Level A and eight Level B Presentation Attack Instrument Species to be used in the evaluation.

Four of eight Level B PAI Species SHALL be tailored to the underlying technology. The PAI Species SHALL be selected by the laboratory and SHALL be approved by the FIDO Biometric Secretariat. Access to the TOE would be necessary. One Presentation Attack Instrument (PAI) SHALL be created for each PAI species and each enrolled test subject.

Note: If the test Laboratory creates PAIs for 8+6=14 selected PAI species and 15 enrolled subjects, the test Laboratory would have to create 210 instruments. Examples of different PAI species include PlayDoh, gelatin using recipe 1, gelatin using recipe 2, and ABC Brand wood glue.

The 15 subjects SHALL have variation in age and gender as well as be representative of the underlying population. Additional target guidance is provided in Population.

The recipes, procedures, and materials to be used for the selected PAI species shall be provided by the Accredited Biometric Laboratory to the vendor well in advance of testing. Recipes SHALL be provided to the FIDO Secretariat. The FIDO Secretariat SHALL ensure that PAI Species selected and created are relatively equivalent between Laboratories.

A check SHALL be performed on each PAI or batch of PAIs to ensure that it is “valid”, i.e., validating that the batch of PAI species captures the biometric characteristic, is prepared properly, and performs as expected. For example, this check could be performed on a test Laboratory reference biometric system that the Laboratory determines is sufficiently similar to the TOE without PAD by observing the biometric image. The Laboratory SHALL document their method for determining that a PAI is valid and submit it as part of the report to FIDO and the Vendor.

If a PAI degrades, additional PAIs SHOULD be created for each enrolled subject.

For each built artifact, the Laboratory shall verify the quality of the artifact before using it in tests. The scope of this quality check is to ensure that the artifact is suitable for verification against the original biometric reference of the test subject.

The quality check can be performed in multiple ways. The particular implementation depends on the biometric modality being tested and on the sensor technology. Existing examples of such quality checks include

• the calculation of an NFIQ2 value for fingerprint images obtained from the artifact

• performing a successful verification with the artifact on a different biometric system without a PAD or with a disabled PAD

6.2.5.1. Impostor Presentation Attack Transactions

For each enrollment, the test Laboratory operator SHALL conduct 10 impostor presentation attack transactions for each PAI. Any kind of guidance SHALL be provided by the biometric authentication system/capture sensor in a similar manner to the final application. The verification transaction (maximum number of verification attempts and timeout period) for a TOE is defined in the Allowed Integration Document as described in Verification Transactions. The transaction SHOULD not exceed 30 seconds. The PAI SHALL be presented the maximum number of attempts allowed for a transaction OR until it matches (which results in an error).

Note: Some presentations may be more successful than others at matching or bypassing PAD. The test crew SHOULD allow for natural variability in presentation of the PAI across the ten transactions.

Note: 15 PAIs will be created for each PAI Species, one for each of 15 enrolled subjects, and 10 imposter presentation attacks transactions will be conducted for each PAI. Therefore, each PAI species will have 150 transactions. To achieve a IAPAR of 7%, there is a maximum of 10 out of 150 errors for each PAI species in order to pass certification. To achieve a IAPAR of 15%, there is a maximum of 22 out of 150 errors for each PAI species in order to pass certification. For IdV Certification only, there is a requirement of less than or equal to 4% for all species, which translates to a maximum of 84 out of 2100 errors.

Note: If it should become clear that a certain kind of material selected for testing is not recognized by the TOE at all (i.e., the TOE does not respond by a rejection, acceptance, or request to try again), the Tester may skip the rest of the impostor transactions for the artifacts of this material. The Lab SHALL test at least three PAIs for that material before determining that it is not recognized by the TOE. This shall be documented in the report. In this case, an IAPAR of 0% shall be recorded as well as the number of PAIS and transactions for the material, and that the testing skipped the remaining PAIs. A failure to acquire for an impostor presentation attack transaction counts as a transaction and does not count as an error, as some systems may produce a failure to acquire in response to a detected presentation attack. In ISO 30107-3, this is considered an attack presentation non-response and computed as the attack presentation non-response rate (APNRR).

6.2.5.1.1. Impostor Presentation Attack Errors

A failure to acquire SHALL be declared when the biometric authentication system is not able to capture and/or generate, biometric features during a verification transaction. The on-line verification test harness SHALL indicate to the Laboratory when a failure to acquire has occurred.

A failure to acquire for an impostor presentation attack transaction counts as a transaction and does not count as an error, as some systems may produce a failure to acquire in response to a detected presentation attack. In ISO 30107-3, this is considered a attack presentation non-response and computed as the attack presentation non-response rate (APNRR).

An impostor presentation attack match error SHALL be declared if the biometric authentication system produces a match decision.

The manner in which the Laboratory records failure to acquire, and impostor presentation attack errors, are left to the laboratory, but SHALL be done automatically to avoid introducing human error.

Note: A verification transaction ends when a decision is made. One or more failures to acquire may occur prior to a decision. The verification transaction (maximum number of verification attempts and timeout period) for a TOE is defined in the Allowed Integration Document as described in Verification Transactions.

6.2.5.1.2. IAPAR

Impostor Attack Presentation Accept Rate (IAPAR) SHALL be calculated according to requirements in Impostor Attack Presentation Accept Rate (IAPAR) .

7. Test Reporting

7.1. General Reporting Requirements

7.1.1. Report Development

The Laboratory SHALL prepare a report as described in the following sections A copy of the report SHALL be provided to the Vendor prior to being provided to FIDO. The report SHALL then be provided to FIDO

The Laboratory SHALL NOT disclose the report to any other recipients; only the Vendor CAN disclose the report to other recipients.

7.1.2. Protection of privacy for test participants

The laboratory report SHALL NOT include the identity and other personal information of the participants.

7.1.3. Description of ToE

Test reports shall provide a description of the ToE. The description shall contain

• complete description of ToE

• ToE configuration files

• ToE settings

• PAD mechanisms

When the ToE is part of a mobile device, the report shall contain the following

• Mobile device model, OS, and OS version

• Position of sensor (e.g. front, back, side), to include position relative to device’s screen(s)

• If applicable, manner of test subject interaction with the biometric sensor (e.g., touch left index finger, swipe right or left thumb, look at front-facing camera, speak a passphrase)

7.1.4. Logging of test activities

In addition to the reports, the Laboratory SHALL maintain a log file in which each interaction (including all attempts from performance testing and all attempts from PAD testing) with the TOE is recorded. The log SHALL include all test attempts, all preparative attempts, management attempts (e.g., setting a threshold) and maintenance activities (e.g., cleaning a sensor). The log SHALL at least contain the following information for each entry:

• Timestamp
• Identity of the tester
• Type of attempt
• Expected outcome
• Actual outcome

The log SHOULD be written automatically by the TOE (cf. requirements for logging for test harness) whenever possible, but may need to be augmented by manual entries that are not known to the TOE. The manual augmentation of the log file is necessary as the TOE does not have the required information to log for some events (e.g., the actual user who performed an impostor attempt under a wrong identity) or will even not be aware of some events (e.g., the fact that a sensor has been cleaned).

The log MUST neither be submitted to FIDO nor the Vendor but remain with the Laboratory. It may be used to answer questions that arise in the context of the certification procedure and is accessible by FIDO upon request.

7.2. FAR/FRR Reporting Requirements

The following SHALL be included in a report to FIDO, following [ISOIEC-19795-1]:

Reporting test details

**Test details**	**Details to report**
The system(s) tested	Including details of algorithms, biometric sensors, user interface, supporting hardware, etc.
Test organization details	Test organization, location, date of test.
Type of evaluation	In the case of technology evaluation: details of the test corpus used. In the case of scenario evaluation: details of the test scenario. In the case of operational evaluation: details of the operational application.
Size of evaluation	Number of test subjects. Number of instances (fingers, hands or eyes, etc.) enrolled by each test subject. Number of visits made by test subject. Number of transactions per test subject (or test subject instance) at each visit.
Test crew	Demographics of the test crew (age, gender, etc.) The manner in which the test crew was assembled, to include exclusions, volunteers etc., as well as the degree to which the test crew mirrored the target population. The level of training, instruction, familiarization, and habituation of test crew in the use of the system.
Test environment	See 8.3.2.1, 8.4.2, and C.2.6.
Time separation between enrollment and recognition transactions	See 7.3.7.
Quality and decision thresholds used during data collection	The thresholds used, and those recommended for the target application (if different).
Control of factors potentially affecting performance	See 7.3 and Annex C.
Test procedures	E.g., policies for determining enrollment failures. Details of any abnormal cases occurring during testing that are excluded from performance analysis.
Estimated uncertainties	Estimated uncertainty in performance results, and method of estimation. See 9.11 and Annex B.
Deviation from guidelines	Deviations from the guidelines of this document should be explained. Sometimes it is necessary to compromise one aspect to achieve another; for example, randomizing the order of using fingers on a fingerprint device might lead to user confusion and a higher number of labelling errors.

Enrollment performance metrics

**Metric**	**Details to report**
Failure to enroll rate (FTER)	See 9.2.1.

Acquisition performance metrics

**Metric**	**Details to report**
Failure-to-acquire rate (FTAR)	See 9.3.1.

Biometric Verification system performance metrics

**Metric**	**Details to report**
False accept rate (FAR)/ False reject rate (FRR)	See 9.5.2 and 9.5.3. FAR and corresponding FRR shall be reported over the range of decision thresholds tested. A DET plot is recommended in the case of multiple operating points.
FTER	See 12.3. Otherwise a statement that FTER is unknown.
FTAR	See 12.4. Otherwise a statement that FTAR is unknown.

Other items of value MAY include:

• Distribution of ethnicity/race
• Additional information as agreed between the laboratory and vendor

7.3. PAD Reporting Requirements

The following SHALL be included in a report to the vendor, following (ISO/IEC 30107-4, 7.2 and 13.4.2.1):

• Summary of the FIDO Biometric Certification and Requirements
• Number of individuals tested
• Distribution of Age
• Distribution of Gender
• Statement relating to selection of the test subjects and the representativeness of the people who will use the system in the real application environment. (This statement does not apply to age and gender which are reported separately.)
• Description of the Test Environment
• Description of the Test Platform
• Number of enrollment transactions
• Number of verification transactions
• Failure to Enroll Rate
• Failure to Acquire Rate
• IAPAR, sample size
• IAPAR of most successful PAI species
• bonafide FRR/FAR calculations and basis of reports; See Section FAR/FRR Reporting Requirements.

And the following from (ISO/IEC 30107-3 and 30107-4):

• Information available to the evaluator about PAD mechanisms in place
• Information about details of any implemented output information of the PAD mechanism other than accept or reject
• Number and description of presentation attack instruments, PAI species, and PAI series used in the evaluation
• number of test subjects involved in the testing
• number of PAI presenters unable to utilize the artifacts
• the purpose and responsibilities of each role in a PAD test, and how the role was material to test results
• number of individuals that assumed each role
• For each role, describe each individuals’ level of experience with presentation attacks
• documentation about the # of individuals that assumed multiple roles (as an incidence matrix)
• information in which machines or automated mechanisms were used as PAI presenters or PAI sources
• number of test subjects unable to present non-conformant characteristics
• number of artifacts created per test subject for each material tested
• number of sources from which the artifact characteristics were derived
• number of tested materials
• Impostor attack presentation accept rate (IAPAR)
• Number of impostor attack presentation transactions
• documentation about ordering of presentations with and without PAIs, and whether PAI presenters or test subjects were reused
• how artifacts were created and prepared (30107-4 10.2)
  • creation and preparation processes
  • effort required to create and prepare artifacts
  • ability to consistently create and prepare artifacts with intended properties
  • customization of artifacts for specific PAI presenters
  • customization of artifacts for specific systems
  • sourcing of biometric characteristics
  • availability of public information on creation and preparation process
  • changes in artifact creation or preparation processes over the course of the evaluation
• how artifacts were used (30107-4 10.3)
  • level of PAI presenter training and habituation
  • artifact durability, including the number of presentations associated with each artifact
  • level of scrutiny or oversight applied during artifact usage
• PAD mechanisms applicable to verification processes (30107-4 11.3)
  • use of quality thresholds and presentation policy
  • parameters of the verification transaction, including the number and duration of presentations
  • level of operator oversight present in the process
  • manner in which operator functions were applied or emulated in the evaluation
  • whether the IUT checks sample quality and provides feedback to the test subject (e.g. “finger too wet”)
  • policy after failing all attempts, e.g., asking for a PIN, a password, or waiting for 30 seconds before attempting again
  • If the IUT provides feedback, a list of the feedback messages

Please note that the log SHALL also include all information about the PAD tests.

Appendix A: Triage of Presentation Attacks by Attack Potential Levels #

For the modalities that can be evaluated by the FIDO test procedures, presentation attacks are described at a high level in Table 1. Table 1 triages presentation attacks into levels based on increasing levels of difficulty to mount and based on frameworks from Common Criteria then applied to biometric presentation attacks in [[FingerprintRecognition], [SOFA-B], [PresentationsAttacksSpoofs], [PAD], [BEAT]]. In ISO 30107-Part 3, this is called the attack potential, defined as the “measure of the capability to attack a TOE given the attacker’s knowledge, proficiency, resources and motivation.”

In [BEAT], the factors are as follows:

1. Elapsed time: <=one day, <=one week, <=one month, >one month
2. Expertise: layman, proficient, expert, multiple experts
3. Knowledge of TOE: public, restricted, sensitive, critical
4. Access to the TOE/Window of Opportunity: easy, moderate, difficult
5. Equipment: standard, specialized, bespoke
6. Access to biometric characteristics: immediate, easy, moderate, difficult

Elapsed time includes time required to create the attack. The definitions for each of the factors are the same as in Section 4.5 in [BEAT].

In [BEAT], these factors are considered for both the Identification and the Exploitation phase. In other words, the factors are scored differently for the phase when the attacker is in the process of identifying the attack compared to the phase where they are actually mounting or exploiting the attack once it has been identified.

For FIDO use case, for Identification phase, we assume that Knowledge of the TOE is “public” and Access to TOE/Window of Opportunity is “easy”, since it would be quite trivial to purchase a sample of the TOE.

1. Knowledge of TOE: public
2. Access to the TOE/Window of Opportunity: easy

Since these factors are generally the same for the majority of FIDO use cases, they are not considered further.

Note: The Window of Opportunity for Biometric Authenticators is impacted by rate-limits on user verification transactions, as required in FIDO Authenticator Security Requirements, Requirement 3.9.

In order to simplify for the FIDO use case, we have collapsed the remaining characteristics into three levels which are described in the next sections. The level rating may change over time as information regarding mountimg an attack will be more broadly disseminated. As such, it is expected that the FIDO Biometric Requirements will be updated in the future to reflect this shift.

The difference of scoring for identification versus exploitation is not considered.

Spoof presentation attack examples separated by levels based on time, expertise, and equipment

		Fingerprint	Face	Iris/Eye	Voice
Level A	Time: < 1 day Expertise: Layman Equipment: Standard	paper printout, direct use of latent print on the scanner	paper printout of face image, mobile device display of face photo	paper printout of face image, mobile device display of face photo	replay of audio recording
	Source of Biometric Characteristic: Immediate, easy	latent fingerprint on the device	photo from social media	photo from social media	recording of voice
Level B	Time: < 7 days Expertise: Proficient Equipment: Standard, Specialized	fingerprints made from artificial materials such as gelatin, silicon.	paper masks, video display of face (with movement and blinking)	video display of an iris (with movement and blinking); printed iris w/ contact lens/doll eye	replay of audio recording of specific pass phrase, voice mimicry
	Source of Biometric Characteristic: Moderate	latent print, stolen fingerprint image	video of subject, high quality photo	video of subject, high quality photo	recording of voice of specific phrase, high quality recording
Level C	Time: > 7 days Expertise: Expert(s) Equipment: Specialized, bespoke	3D printed spoofs	silicon masks, theatrical masks	contact lens/prosthetic eye with a specific pattern	voice synthesizer
	Source of Biometric Characteristic: Difficult	3D fingerprint information from subject	high quality photo, 3D face information from subject	high quality photo in Near IR	multiple recordings of voice to train synthesizer

Level A

Level A attacks are quite simple to carry out and require relatively little time, expertise, or equipment. Biometric characteristics under attack are quite easy to obtain (e.g. face image from social media, fingerprint from the device and reused directly).

1. Elapsed time: <=one day
2. Expertise: Layman
3. Equipment: Standard
4. Access to biometric characteristics: Immediate, easy

Level B

Level B attacks require more time, expertise and equipment. Additionally, the difficulty to acquire the biometric characteristic is higher (e.g., stolen fingerprint image, high quality video of a person’s face).

1. Elapsed time: <=one week
2. Expertise: Proficient
3. Equipment: Standard, Specialized
4. Access to biometric characteristics: Moderate

If at least one of these characteristics reaches the levels listed above, the attack is categorized as Level B.

Level C

Level C includes the most difficult attacks.

1. Elapsed time: <=one month, >one month
2. Expertise: Expert, multiple experts
3. Equipment: Specialized, bespoke
4. Access to biometric characteristics: Difficult

If at least one of these characteristics reaches the levels listed above, the attack is categorized as Level C.

PAI Species for Fingerprint

Level A attacks for spoofing a fingerprint biometric are to retrieve and print an image of a fingerprint which can be obtained through taking an image of a dusted latent fingerprint. This requires equipment that is readily available to most individuals and requires very little skill. Examples of different Presentation Attack Instrument Species (PAI Species) for Level A are a set of fingerprint images printed on inkjet printers or laser printers. Each make/model of the printer would be considered a species. In addition, preprocessing could be used to enhance the image. Any alterations such as this would be categorized as a different PAI species. Some TOEs may be based on a photograph of a person’s finger(s). Level A spoof attacks for this type of fingerprint TOE could include a low resolution photograph of a person’s hand. In attacks of this type, photographs may happen to include a hand, but are likely to be of low resolution.

Level B attacks for spoofing a fingerprint biometric are to retrieve and print an image of a fingerprint which can be obtained through taking an image of a dusted latent fingerprint or retrieving a stolen fingerprint image from a database or other source of stolen fingerprint images. Level B attacks are similar to Level A attacks, except rather than simply printing a fingerprint image, the image could be converted into a mold. A mold could be created through etching a printed circuit board, laser etching, or simply printing a fingerprint image on a transparency. A 2D mold can also be made using an inexpensive 3D printer (e.g. less than $500) with sufficient resolution to print the fingerprint ridges (e.g. at a minimum, XY resolution of less than 0.05mm). Once a mold is created, a PAI (or cast) can be created by placing other materials such as gelatin, silicon, play-doh, etc into the mold. The difficulty of making the translation from a 2D fingerprint image to a mold moves this attack from Level A to Level B. Additives could be added to the PAI to increase conductivity such as graphite or lotion. Any alterations such as this would be categorized as a different PAI species. Some TOEs may be based on a photograph of a person’s finger(s). In additional to attacks based on materials like gelatin or PlayDoh, Level B spoof attacks for this type of fingerprint TOE could include a high resolution photograph or a video of a person’s hand.

Level C attacks are more elaborate and capture additional information such as pores, veins, sweating, and 3D details. PAI could also be a 3D printed finger. Some molds may also be more elaborate, such as 3D printing. Level C PAIs take more time to create, are more expensive, require experts to prepare, and need a high resolution and/or 3D finger information.

Note: For creation of Presentation Attack Instruments (PAI) during testing, test subjects will provide biometric characteristics on which the PAI will be based through pressing their finger on a surface creating a latent print (Level A and above), taking a low-resolution photograph (Level A), capturing their fingerprint on a fingerprint scanner (Level B and above), or taking a high-resolution photograph (Level B and above). Fingerprint molds obtained from an individual through pressing a finger into silicon or other molding material are out of scope. Future PAD testing may include molds of fingerprint when attacks of this type impact FIDO-based use cases.

Table of Example PAI Species for Fingerprint

Species	Level
latent fingerprint captured and printed on inkjet or laser printer	A
Low resolution photograph of a person’s fingers	A
Fingerprint molds created using PCB or laser etching	B
Fingerprint molds created by printing on a transparency	B
Fingerprint molds created by printing with an inexpensive (e.g. less $500) 3D printer with XY resolution of less than 0.05mm	B
Casts made from molds listed above with materials such as gelatin or silicone	B
Same as previous with graphite or other material placed on surface of mold or PAIs	B
High resolution photograph of a person’s fingers	B
Laser etching a fingerprint directly on materials such as rubber or silicone	B
Video (low or high resolution) of a person’s fingers	B
3D printed molds and/or PAIs with expensive, high resolution 3D printers	C
Fingerprint models which capture sweating, veins, blood flow or more sophisticated finger information	C

Note: Some TOEs may involve multiple fingerprints. PAIs should be created for each finger that is used in making a decision.

PAI Species for Face

Level A attacks for spoofing a face biometric are to retrieve and utilize a photograph of the individual under attack. For example, an attacker can copy a photograph from a social media site and print the photograph or display the photo on a mobile device to the biometric recognition system. This requires equipment that is readily available to most individuals and requires very little skill. Examples of different Presentation Attack Instrument Species (PAI Species) for Level A are a set of face images printed on inkjet printers, laser printers, or photographs printed at a photograph laboratory. Each make/model of the printer would be considered a species. In addition, preprocessing could be used to enhance a photograph, as well as holes could be cut out for the eyes, nose, mouth, or outline of face. Any alterations such as this would be categorized as a different PAI species. Examples of different PAI species for displayed photos on mobile devices would be changing a phone, tablet, and computer monitor make/models. Level A also includes videos created by readily available, inexpensive deepfake tools which can animate a face based on a single photograph of an individual. Animation of a 2D image of a person’s face may include blinking, smiling, or speaking. The videos are then displayed on electronic/mobile devices and used to attack a face recognition system. Injection attacks are out of scope for PAD testing.

Note: Some face recognition systems utilize NIR illumination and a NIR camera. However, photographs of an individual taken in the visible spectrum could be used as a source of the face biometric characteristics for creation of a PAI. Preprocessing the RGB image may be needed, such as only selecting the red channel. NIR reflectance of the printed photograph may also be impacted by the make and model of the printer used.

Note: Some face recognition systems utilize NIR illumination and a NIR camera for which mobile display of face may not be feasible for most device makes and models.

Level B attacks are similar to Level A attacks, except rather than a face photograph, a video of the subject is needed. The difficulty in acquiring a video rather than a photograph is what moves it from Level A to Level B; even though it is possible, it is less likely to obtain a video of a person compared to a photograph. Additionally, with a high resolution face image, it is possible to create a paper mask of the person. This requires proficient expertise and therefore is also included as a Level B attack. Examples of different PAI species for displayed videos on electronic devices would be changing a phone, tablet, and computer monitor make/models. Level B also includes videos created by readily available, inexpensive deepfake tools which can animate a face based on multiple and/or video frames of an individual. Animation of a person’s face may include blinking, smiling, or speaking. The videos are then displayed on electronic/mobile devices and used to attack a face recognition system.

Note: As with Level A, some face recognition systems utilize NIR illumination and a NIR camera for which mobile display of face or face video may not be feasible for most device make and models.

Level C attacks involve more elaborate masks that are not made of paper, but rather other specialized materials (e.g. ceramic, silicone). These masks take more time to create, are more expensive, and need a high resolution photograph and/or 3D information. 3D information can also be derived from a 2D photo using sophisticated computer vision techniques. Masks include rigid 3D with and without eye holes, flexible silicone masks, and 3D printed, color face replicas. A video can also be created by using sophisticated computer vision which animates a 2D image of a person’s face to blink, smile, or speak (e.g. DeepFake). Level C also includes videos created by more sophisticated deepfake tools which can animate a face based on multiple and/or video frames of an individual. Animation of a person’s face may include blinking, smiling, or speaking. The videos are then displayed on electronic/mobile devices and used to attack a face recognition system.

Note: Twins or genetically identical siblings may be more likely to have a similar face signature. We have not included twins or genetically identical siblings in the attacks that are being tested.

Table of Example PAI Species for Face

Species	Level
Face image printed on inkjet or laser printer	A
Face image printed at photograph laboratory	A
Displayed photos on electronic/mobile devices	A
Videos created by readily available, inexpensive deepfake tools which can animate a face based on a single photograph of an individual (displayed on electronic/mobile devices)	A
Displayed videos on electronic/mobile devices	B
Paper masks	B
Videos created by readily available, inexpensive deepfake tools which can animate a face based on multiple and/or video frames of an individual (displayed on electronic/mobile devices)	B
Masks made of specialized materials (ceramic, silicone, and/or theatrical)	C
3D printed faces	C
Videos created by more sophisticated deepfake tools which can animate a face based on multiple and/or video frames of an individual (displayed on electronic/mobile devices)	C

PAI Species for Iris/Eye

Level A attacks for spoofing an iris or eye biometric are to retrieve and utilize a photograph of the individual under attack. For example, an attacker can copy a photograph from a social media site and print the photograph or display the photo on a mobile device to the biometric recognition system. This requires equipment that is readily available to most individuals and requires very little skill. Examples of different Presentation Attack Instrument Species (PAI Species) for Level A are a set of images of an iris or eye printed on inkjet printers, laser printers, or photographs printed at a photograph laboratory. Each make/model of the printer would be considered a species. In addition, preprocessing could be used to enhance a photograph, as well as holes could be cut out for the pupils. Any alterations such as this would be categorized as a different PAI species. Examples of different PAI species for displayed photos on mobile devices would be changing a phone, tablet, and computer monitor make/models.

Note: A majority of iris systems utilize NIR illumination and a NIR camera. For example, in typical visible spectrum images, the iris pattern does not show for dark eyes, but may be visible for light colored eyes (e.g. blue). Thus, photographs of an individual taken in the visible spectrum may not be an effective source of the iris biometric characteristics and needs to be considered when constructing an attack. Preprocessing the RGB such as only selecting the red channel may be needed.

Note: A majority of iris systems utilize NIR illumination and a NIR camera for which mobile display of iris may not be feasible for most device make and models.

Level B attacks are similar to Level A attacks, except rather than a photograph, a video of the subject is needed. The difficulty in acquiring a video rather than a photograph is what moves it from Level A to Level B; even though it is possible, it is assumed to be more difficult to obtain a video of a person compared to a photograph. This requires proficient expertise and therefore is also included as a Level B attack. Examples of different PAI species for displayed videos on electronic devices would be changing a phone, tablet, and computer monitor make/models. Another example of a Level B attack is inserting a printed iris into a fake eye that is readily available, e.g. doll eye. A printed eye with a contact lens on top is another example of a Level B attack.

Note: As with Level A, a majority of iris systems utilize NIR illumination and a NIR camera for which mobile display of iris may not be feasible for most device make and models.

Level C attacks involve more elaborate eye prosthetics that are not made of paper, but rather silicon or materials that have similar spectral characteristics as a human eye and/or iris. These prosthetics take more time to create, are more expensive, and need a high resolution photograph, 3D information, and/or spectral characteristics of the eye and/or iris.

Table of Example PAI Species for Iris/Eye

Species	Level
Iris/eye image printed on inkjet or laser printer	A
Iris/eye image printed at photograph laboratory	A
Displayed Iris/eye photos on electronic/mobile devices	A
Displayed Iris/eye videos on electronic/mobile devices	B
Printed iris/eye inserted in fake eye	B
Printed eye with contact lens on top	B
Prosthetic eye	C
Prosthetic eye with similar spectral characteristics to human eye	C

PAI Species for Voice

Level A attacks for spoofing a voice biometric are to retrieve and utilize a voice recording of the individual under attack. For example, an attacker can record the voice of the individual and replay their voice to the biometric recognition system. This requires equipment that is readily available to most individuals and requires very little skill. Examples of different Presentation Attack Instrument Species (PAI Species) for Level A are a set of voice recordings replayed on different speakers. Each make/model of the recording device and speakers to replay the voice would be considered a species. In addition, preprocessing could be used to enhance the audio. Any alterations such as this would be categorized as a different PAI species. Equipment used for preprocessing, recording, and replay should be standard, readily available, easy to use equipment.

Level B attacks are similar to Level A attacks, except rather than any voice recording, a recording of a specific passphrase is needed. The difficulty in acquiring a recording of specific set of words rather than any words is what moves it from Level A to Level B; even though it is possible, it is less likely to obtain a recording of a specific passphrase. Also, multiple speech recordings from a person could be used to attack the system by cutting portions of words needed in a phrase using commodity off the shelf audio editors. Additionally, high quality recording and replay equipment also would be considered a Level B attack, where each equipment set-up would be considered a different PAI. Leval B attacks also include readily available voice synthesizers which can take a recording of a voice, build a model of that voice, and replay a person speaking any words.

Note: Text-independent systems may be vulnerable to any recording, where text-dependent systems are vulnerable to a recording of the specific pass-phrase.

Level C attacks involve more sophisticated voice synthesizers which are built using speech samples from a large population, then tuned for a specific individual. These models take more time to create, require more skill, and need high resolution, long recordings to build accurate models. A person may also be skilled in the art of impersonation where they attempt to mimic someone else.

Note: Twins or genetically identical siblings may be more likely to have a similar voice signature. We have not included twins or genetically identical siblings in the attacks that are being tested.

Table of Example PAI Species for Voice

Species	Level
Recording a voice saying any words from readily available equipment for recording and playback	A
Recording a voice saying specific passphrase from readily available equipment for recording and playback	B
Recordings of a specific passphrase created by cutting and pasting together words using readily available software	B
High quality recording a voice saying any words from high end equipment for recording and playback	B
Readily available, inexpensive voice synthesizers which can be trained based on short recordings of an individual and playback any words	B
More sophisticated voice synthesizer which can playback any words, trained from long, high quality recordings or a database of recordings	C
Impersonation, where an attacker is able to mimic a person’s voice	C