In-distribution adversarial attacks on object recognition models using gradient-free search
This reveals a worrisome problem for AI safety as in-distribution adversarial examples bypass the need for engineered noise, potentially affecting all object recognition models.
The paper challenges the claim that adversarial failures are due to out-of-distribution data by demonstrating that perturbed images within the training distribution can cause misclassifications, with CMA-Search finding failures in over 71% of cases for camera position changes and 42% for lighting changes.
Neural networks are susceptible to small perturbations in the form of 2D rotations and shifts, image crops, and even changes in object colors. Past works attribute these errors to dataset bias, claiming that models fail on these perturbed samples as they do not belong to the training data distribution. Here, we challenge this claim and present evidence of the widespread existence of perturbed images within the training data distribution, which networks fail to classify. We train models on data sampled from parametric distributions, then search inside this data distribution to find such in-distribution adversarial examples. This is done using our gradient-free evolution strategies (ES) based approach which we call CMA-Search. Despite training with a large-scale (0.5 million images), unbiased dataset of camera and light variations, CMA-Search can find a failure inside the data distribution in over 71% cases by perturbing the camera position. With lighting changes, CMA-Search finds misclassifications in 42% cases. These findings also extend to natural images from ImageNet and Co3D datasets. This phenomenon of in-distribution images presents a highly worrisome problem for artificial intelligence -- they bypass the need for a malicious agent to add engineered noise to induce an adversarial attack. All code, datasets, and demos are available at https://github.com/Spandan-Madan/in_distribution_adversarial_examples.