Continual learning worsens fairness issues in deepfake detection over time
Time to Fake It: Continual Machine Learning Intensifies Unfairness in Deepfake Detection
Machine LearningArtificial Intelligence
Summary
Detecting deepfake videos is important to stop misinformation, but methods that update over time to catch new fakes can become less fair to different groups of people. The authors studied how these detection systems change as they learn new tricks and found that their accuracy and fairness often get worse together. They also discovered that the ways these systems forget old information can make some groups less protected, not just because those groups have less data, but due to how the system learns overall. This work suggests that improving fairness should be a key part of ongoing training for deepfake detectors.
What this means in practice
- •For security software teams: Design deepfake detection systems that consider fairness degradation when updating models with new data.
- •For ai ethics compliance teams: Evaluate ongoing impacts on fairness of AI systems that adapt to evolving content like deepfakes across diverse user groups.
Authors
Kajaree Das, Oliver Böse, Subarnaduti Paul, Gabriel Storch, Martin Mundt
Abstract
Deepfake detectors safeguard against misinformation, rendering their failures as much a societal as a technical concern. As generators constantly evolve, detectors trained on static datasets miss newly arising threats, while progressive training interferes catastrophically. Yet, mitigating forgetfulness in terms of accuracy, the prevalent focus of continual learning (CL), is insufficient to reflect real-world impact. We therefore scrutinize the effects of CL methods with respect to evolving fairness. To this end, we analyze four dimensions of fairness across diverse demographics for detectors trained progressively on a chronological sequence of 11 real-world deepfake generation methods. Tracking fairness metrics over time, we expose non-trivial interplays between performance and fairness that depend on both the choice of CL algorithm and demographic intersections. We find accuracy degradation to typically exacerbate unfairness and, conversely, fairness improvements to remain scarce over time. Dissecting the CL method families (rehearsal, prompts, regularization), we observe that inequitable forgetting is not explained by underrepresentation alone, but likely entangled with representation learning. We thus issue a call to action for CL methods to take social outcomes into account when balancing the stability-plasticity tradeoff.