Ido Nitzan Hidekel

Neural networks preferentially learn low frequencies, which both leaves high-frequency artifacts in generated audio and leaves those artifacts under-exploited by common detectors. SONAR is a dual-path framework that learns data-driven SRM-style filters to isolate high-frequency residuals and uses a Jensen–Shannon contrastive objective to pull real content/noise pairs together and push fake pairs apart in latent space, yielding strong generalization to unseen generators.

@inproceedings{hidekel2026sonar,
  title     = {SONAR: Spectral-Contrastive Audio Residuals for Generalizable Deepfake Detection},
  author    = {Hidekel, Ido Nitzan and Lifshitz, Gal and Cohen, Khen and Raviv, Dan},
  booktitle = {Proceedings of the International Conference on Machine Learning (ICML)},
  year      = {2026}
}

A decade of mitigation strategies later, no theory explains why catastrophic forgetting happens. We give a function-space account in the Neural Tangent Kernel regime: new-task training induces interference along a structured, low-rank subspace of output space defined by the cross-task NTK. A closed-form predictor identifies the forgetting vector (direction and magnitude jointly) before new-task training begins, and the theory shows forgetting concentrates in just 1–6 NTK eigenmodes, explaining why parameter-space methods fail on shared-head benchmarks and motivating a targeted spectral regularizer.

@inproceedings{hidekel2026forgetting,
  title     = {Catastrophic Forgetting is Low-Rank: A Function-Space Theory for Continual Adaptation},
  author    = {Hidekel, Ido Nitzan and Raviv, Dan},
  booktitle = {ICML 2026 Workshop on Continual Adaptation at Scale: Towards Sustainable AI (CATS)},
  year      = {2026}
}