Quantifying Bias in a Face Verification System
Machine learning models perform face verification (FV) for a variety of highly consequential applications, such as biometric authentication, face identification, and surveillance. Many state-of-the-art FV systems suffer from unequal performance across demographic groups, which is commonly overlooked...
| Auteurs principaux: | , , , , |
|---|---|
| Autres auteurs: | |
| Format: | Article |
| Publié: |
Multidisciplinary Digital Publishing Institute
2022
|
| Accès en ligne: | https://hdl.handle.net/1721.1/142034 |
| _version_ | 1826203055227404288 |
|---|---|
| author | Frisella, Megan Khorrami, Pooya Matterer, Jason Kratkiewicz, Kendra Torres-Carrasquillo, Pedro |
| author2 | Lincoln Laboratory |
| author_facet | Lincoln Laboratory Frisella, Megan Khorrami, Pooya Matterer, Jason Kratkiewicz, Kendra Torres-Carrasquillo, Pedro |
| author_sort | Frisella, Megan |
| collection | MIT |
| description | Machine learning models perform face verification (FV) for a variety of highly consequential applications, such as biometric authentication, face identification, and surveillance. Many state-of-the-art FV systems suffer from unequal performance across demographic groups, which is commonly overlooked by evaluation measures that do not assess population-specific performance. Deployed systems with bias may result in serious harm against individuals or groups who experience underperformance. We explore several fairness definitions and metrics, attempting to quantify bias in Google’s FaceNet model. In addition to statistical fairness metrics, we analyze clustered face embeddings produced by the FV model. We link well-clustered embeddings (well-defined, dense clusters) for a demographic group to biased model performance against that group. We present the intuition that FV systems underperform on protected demographic groups because they are less sensitive to differences between features within those groups, as evidenced by clustered embeddings. We show how this performance discrepancy results from a combination of representation and aggregation bias. |
| first_indexed | 2024-09-23T12:30:40Z |
| format | Article |
| id | mit-1721.1/142034 |
| institution | Massachusetts Institute of Technology |
| last_indexed | 2024-09-23T12:30:40Z |
| publishDate | 2022 |
| publisher | Multidisciplinary Digital Publishing Institute |
| record_format | dspace |
| spelling | mit-1721.1/1420342023-02-14T19:43:47Z Quantifying Bias in a Face Verification System Frisella, Megan Khorrami, Pooya Matterer, Jason Kratkiewicz, Kendra Torres-Carrasquillo, Pedro Lincoln Laboratory Machine learning models perform face verification (FV) for a variety of highly consequential applications, such as biometric authentication, face identification, and surveillance. Many state-of-the-art FV systems suffer from unequal performance across demographic groups, which is commonly overlooked by evaluation measures that do not assess population-specific performance. Deployed systems with bias may result in serious harm against individuals or groups who experience underperformance. We explore several fairness definitions and metrics, attempting to quantify bias in Google’s FaceNet model. In addition to statistical fairness metrics, we analyze clustered face embeddings produced by the FV model. We link well-clustered embeddings (well-defined, dense clusters) for a demographic group to biased model performance against that group. We present the intuition that FV systems underperform on protected demographic groups because they are less sensitive to differences between features within those groups, as evidenced by clustered embeddings. We show how this performance discrepancy results from a combination of representation and aggregation bias. 2022-04-25T12:36:10Z 2022-04-25T12:36:10Z 2022-04-20 2022-04-21T21:03:49Z Article http://purl.org/eprint/type/JournalArticle https://hdl.handle.net/1721.1/142034 Computer Sciences & Mathematics Forum 3 (1): 6 (2022) PUBLISHER_CC http://dx.doi.org/10.3390/cmsf2022003006 Creative Commons Attribution https://creativecommons.org/licenses/by/4.0/ application/pdf Multidisciplinary Digital Publishing Institute Multidisciplinary Digital Publishing Institute |
| spellingShingle | Frisella, Megan Khorrami, Pooya Matterer, Jason Kratkiewicz, Kendra Torres-Carrasquillo, Pedro Quantifying Bias in a Face Verification System |
| title | Quantifying Bias in a Face Verification System |
| title_full | Quantifying Bias in a Face Verification System |
| title_fullStr | Quantifying Bias in a Face Verification System |
| title_full_unstemmed | Quantifying Bias in a Face Verification System |
| title_short | Quantifying Bias in a Face Verification System |
| title_sort | quantifying bias in a face verification system |
| url | https://hdl.handle.net/1721.1/142034 |
| work_keys_str_mv | AT frisellamegan quantifyingbiasinafaceverificationsystem AT khorramipooya quantifyingbiasinafaceverificationsystem AT mattererjason quantifyingbiasinafaceverificationsystem AT kratkiewiczkendra quantifyingbiasinafaceverificationsystem AT torrescarrasquillopedro quantifyingbiasinafaceverificationsystem |