Antifragile Deepfake Detection: Robustness, Adversarial Threats, and Open-Set Challenges for Digital Evidence
The rapid progress of generative AI is making deepfakes increasingly realistic, accessible, and difficult to assess in uncontrolled online environments. This talk addresses the challenge of detecting deepfakes “in the wild”, where multimedia content is often compressed, shared across platforms, generated by unseen models, or deliberately modified through adversarial attacks. Particular attention will be given to the limits of current detectors in terms of generalization and out-of-distribution robustness, highlighting recent approaches based on continual learning, OOD detection, energy-based scoring, uncertainty-aware methods and open-set recognition for digital evidence scenarios. The lecture will also discuss the role of multimodal analysis, integrating visual, audio, textual, and contextual cues to improve reliability beyond single-modality detectors. Recent directions involving large language models and AI agents will be considered as emerging tools for misinformation analysis, evidence interpretation, and forensic support. Finally, the talk will emphasize the importance of explainability and mechanistic interpretability, showing how XAI methods can help analysts interpret model decisions, identify failure cases, and increase trust in deepfake detection systems used in forensic and investigative contexts.