Adversarial perturbations crafted on public ASR models transfer to their privately fine-tuned versions with near white-box strength. TransferBreaker suppresses this transfer during fine-tuning, combining base-model adversarial training, latent Jacobian regularization, and hybrid base/target perturbations. Across three languages and four ASR models, it cuts adversarial WER from 92.6 to 27.8.

Unlike prior phrase-level ASR backdoors, GhostWord is a word-level, composable attack. A codebook maps imperceptible audio patterns to target words, each overlaid on a force-aligned spoken word with a matching transcript swap. It reaches 89.3% success and transfers across models and languages. Defenses that suppress it roughly double clean WER.

Weakly supervised video anomaly detection models are vulnerable to adversarial attacks, and standard defenses fail under weak supervision. FrameShield uses SRD, a pseudo-anomaly generator that creates temporally consistent synthetic anomalies, to reduce pseudo-label noise and enable effective adversarial training, outperforming prior methods by 71.0% AUROC on average.

PatchGuard is an adversarially robust anomaly detection and localization method built on Vision Transformers. It trains with foreground-aware pseudo-anomalies and a novel loss, achieving large robustness gains on industrial and medical datasets.

Anomaly detectors trained only on normal samples are vulnerable to adversarial attacks. We build a pseudo-anomaly group and apply adversarial training with a contrastive loss, using opposite pairs to mitigate spurious negative pairs and improve robustness.

Novelty detection methods are usually tailored to a dataset's inductive biases and fail to generalize. We propose a universal novelty detector that adapts to diverse datasets through adaptive contrastive learning.

We detect whether a model is trojaned, based on the finding that backdoored models exhibit jagged decision boundaries around out-of-distribution samples, which reduces their robustness there.