Adversarial Example Detection
Overview
Adversarial example detection studies how to distinguish deliberately perturbed inputs from benign ones when small changes can cause a model to make errors. Methods look for signals such as unusual input frequencies, shifts in internal representations, prediction instability under transformations, or disagreement across layers and tasks; some use uncertainty estimates or formal certification. These approaches are studied in image, speech, language, and multimodal systems, with detectors ranging from training-free checks to learned classifiers calibrated on benign data. A central challenge is detecting adaptive and previously unseen attacks while keeping false alarms and computational costs low, and empirical success does not always provide a formal guarantee.
Latest papers
50 papers3 papers in the last four weeks, down 70% on the four weeks before. 0.0% of all new papers.