Preliminary User Evaluation of Deepfake Detection System in Criminal Justice Facial Evidence Verification

�IPSJ 86^th Conference Presentation 2024

Ebrima Hydara, Masato Kikuchi, Tadachika Ozono

Nagoya Institute of Technology | 21 February 2024

​

​

1

Introduction: Research Goal

Problem

• Deepfake: highly realistic fake media
• Facial evidence: proof by facial identity
• Current facial evidence verification methods cannot tackle deepfake evidence (Wehrli et al. 2022)

​

Solution

• User-centric deepfake detection system for facial evidence veracity check
• Use human-in-the-loop method for improvement
• Incorporate forensic techniques to comply with judicial proceedings
• Safeguard evidence integrity in criminal justice

2

Related Work: Current Methods

Facial Evidence Verification:

• Face-matching (Bacci et al., 2021), (Moreton, 2021)
• Facial Recognition Technology (Hill, D. et al., 2022)
• Automated Facial Recognition System (Khan et al., 2021)
• They all use face matching as main technique

Deepfake Detection:

• Threshold Classifier (Reis & Ribeiro, 2023)
• Likelihood Ratio Framework: (Meuwly, Ramos, & Haraksim, 201)
• Statistical Analysis: (Morrison, 2022)
• Biometric-Based Forensic Technique: (Agarwal et al., 2020)
• CNN, RCNN, & ViT-based architectures (Silver et al., 2022)

3

Related Work: Gap Analysis

Current Deepfake Detection:

• No decision-making explainability for criminal justice
• No confidence boosting mechanism for forensics
• No individual frame prediction, timestamp & heatmap
• No specialized reporting scheme for criminal justice

Proposed System Method:

• Synergy of deepfake detection & forensic techniques to explain decision-making
• Confidence boosting for forensics
• Individual frame prediction, timestamp, heatmap
• Specialized reporting for criminal justice

4

Contribution: Originality/Novelty

• Method: the combination of the following to holistically address the problem:
• Deepfake Detection to detect veracity of evidence
• Confidence Threshold to filter only trustworthy predictions, minimizing false positives, ensuring the credibility of evidence, and preserving the judicial proceeding’s integrity
• Frame Timestamps to track events and enable verification of temporal integrity of evidence
• Frame Heatmap to visualize regions of manipulations for forensics

5

Methodology: Training & Evaluation

6

Training & Testing

Accuracy > 80%

Datasets

System Implementation: Architecture

7

System Implementation: Prototype

8

System Implementation: Demo

9

Experiments: Setup

• Objective: measure system-user collaborative performance & user confidence on system’s performance
• Dataset: 200 videos of both deepfake and authentic footage
• Participants: 10 non-expert users from varying backgrounds
• Records: standardized recording of results in Excel
• Analysis: accuracy, precision, recall, & F1 score measurement

10

Experiments: Results (System Performance)

11

Best Performance Confidence Threshold:

80 to 85

Experiments: Results

• Collaborative performance evaluation results:
• User perceptions resonate with system predictions
• Users were more careful to misrepresent footage as deepfake due to their human resilience

12

Experiments: Results

• User confidence spectrum in system’s predictive performance results:
• Most users are confident in the system’s predictions
• User confidence should however be strengthened around recall (minimize FP)

13

Discussion

Strength

• Accuracy & Precision: innocent suspects can be acquitted of false charges by just 1 frame predicted as fake
• User confidence: promising outlook for the criminal justice domain where confidence & trust are required

Limitations

• Recall: lower system recall than human judgment recall calls for improvement on system’s resilience
• No criminal justice deepfake datasets
• No access to forensic experts for evaluations

14

Conclusion: Findings & Future Work

Findings

• Accuracy and Precision scores show high collaborative user-system performance, however, Recall and F1 scores show higher human resilience than the system
• A user-centric approach is needed for a highly performant deepfake detection system for facial evidence verification in criminal justice.

Future Work

• Better explainability and feedback
• Better system-user interaction transparency
• Improve system’s recall for more accountability and trust
• Engage Forensic experts for system evaluation
• Define new standard operating procedures for policy guidelines

15

References

• Bacci, N. et al.: Forensic Facial Comparison: Current Status, Limitations, and Future Directions. Biology, vol. 10, no. 12, 1269, pp. 1–26 (2021).
• Moreton, R.: Forensic face matching: Procedures and application. In M. Bindemann (Ed.), Forensic face matching: Research and practice, pp. 144–173 (2021).
• Hill, D. et al.: Police use of facial recognition technology: The potential for engaging the public through co-constructed policy-making. International Journal of Police Science and Management, vol. 24, no. 3, pp. 325–335 (2022).
• Khan, Z. A. et al.: AI-Based Facial Recognition Technology and Criminal Justice: Issues and Challenges. Turkish Journal of Computer and Mathematics Education, vol. 12, no. 14, pp. 3384–3392 (2021).
• Dosovitskiy, A. et al.: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In: International Conference on Learning Representations (2021).
• Towler, A. et al.: Unfamiliar face matching systems in applied settings. In: pp. 21–40, (2017).
• Wehrli, S. et al.: Bias, awareness, and ignorance in deep-learning-based face recognition. In: AI Ethics 2, pp. 509–522, (2022).

​

​

16

References

• Reis, P. et al. (2023). “A forensic evaluation method for DeepFake detection using DCNN-based facial similarity scores.” In: Forensic science international, p. 111747. URL: https://api.semanticscholar.org/CorpusID:259070218.
• Meuwly, D. et al. (2017). “A guideline for the validation of likelihood ratio methods used for forensic evidence evaluation.” In: Forensic science international 276, pp. 142–153. URL: https://api.semanticscholar.org/CorpusID:4785352.
• Morrison, G. (2022). “Advancing a paradigm shift in evaluation of forensic evidence: The rise of forensic data science”. In: Forensic Science International: Synergy 5, p. 100270. DOI: 10 . 1016 / j . fsisyn . 2022 . 100270.
• Agarwal, S. et al. (2020). “Detecting deep-fake videos from appearance and behavior”. In: 2020 IEEE International Workshop on Information Forensics and Security (WIFS). IEEE, pp. 1–6.
• Hu, J. et al. (2021). “Improving the generalization ability of deepfake detection via disentangled representation learning”. In: 2021 IEEE International Conference on Image Processing (ICIP). IEEE, pp. 3577–3581.
• Le, B. M. et al. (2023). “Why do deepfake detectors fail?” In: arXiv. arXiv: 2302.13156 [abs].
• Akhtar, Z. (2023). “Deepfakes Generation and Detection: A Short Survey”. In: Journal of Imaging 9.1, p. 18. DOI: 10.3390/jimaging9010018.

​

17

Thank You

18