Adversarial examples beyond spurious features: An experimental study

Summary

This study examines how spurious correlations, non-causal links between superficial attributes and labels, create adversarial vulnerabilities in neural networks. The authors propose two spurious-free gradient-based evasion attacks built on assumptions of biased gradients toward simple features and overconfident output probabilities, and evaluate them on benchmark vision datasets against standard and spurious-robust models. Results show the attacks diminish the influence of spurious attributes, with each performing best in specific scenarios.

Topics

• Adversarial machine learning

Related items

• InfoImage watermarking robustness under MLLM-guided semantic editing: Benchmark, defense, and diagnosisSimilar attack · Elsevier Security Journals
• LowYour Robot's Safety Functions Already Work. What If the Input Lies?Similar attack · VicOne Blog
• InfoSpeech Tokenizers are Vulnerable: Transferable Semantic Attack and Robust TokenizerSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)
• InfoAdversarial Retraining for Machine Learning-Based Man-in-the-Middle Attack Detection in Smart Grid Industrial Control SystemsSimilar attack · OpenAlex (peer-reviewed AI security)
• LowMeasuring Robustness and Efficiency in a Connectome-Constrained Fly Visual System Model: Linking an Adversarial and a Perceptual IllusionSimilar attack · Arxiv (cs.RO + cs.CV security)