Abstract

Generative AI has shattered one of our most fundamental assumptions — that seeing is believing. Today, hyper-realistic deepfakes are weaponizing trust across finance, healthcare, law, media, and critical infrastructure, enabling synthetic identity fraud, voice-cloned authorizations, forged legal evidence, and biometric spoofing at unprecedented scale. Existing detection tools are losing the arms race. This talk argues that the solution is not faster detectors — it is trustworthy ones. The talk opens by defining the core properties of a trustworthy deepfake detector and introducing a Neurosymbolic AI framework that fuses deep neural perception with symbolic reasoning, producing systems that don't merely flag suspicious media but can explain why, in auditable, human-readable logic. From there, the talk addresses four critical real-world challenges that expose the limits of conventional approaches. First, zero-day attacks: rather than memorizing known forgery signatures, robust detectors must generalize to entirely unseen generative models through manipulation-agnostic feature learning and open-set recognition. Second, partial fake detection: real-world forgeries increasingly involve selective manipulation,  a swapped face in isolated frames or a spliced audio segment, demanding localization-aware architectures that pinpoint tampered regions rather than rendering binary verdicts. Third, compression robustness: deepfakes in the wild survive multiple rounds of lossy compression that strip away forensic artifacts, requiring frequency-domain analysis and compression-aware training to maintain detection integrity across real-world distribution pipelines. The talk then broadens the definition of trustworthiness beyond accuracy to encompass three pillars: explainability and source tracing, which links detected fakes back to their generative origin for forensic and legal accountability; fairness and diversity, addressing dangerous demographic blind spots that cause detectors to underperform; and robustness, ensuring reliable performance under distribution shift and adaptive adversaries. Finally, the talk presents defense mechanisms against adversarial attacks.