Gallery for "Exploring Adversarial Robustness of Deep Metric Learning"

Examples of inference across various DML models (both naturally-trained and robust) is shown below. Benign (unperturbed) input is denoted as \(x\), while the adversarial perturbed variant is denoted as \(x'\). The function \(\text{nn}(\cdot)\) outputs the nearest anchor in the embedding space. Green borders reflect correct inference (anchor has same class), while red reflect incorrect inference. Recall, that \(x\) is only perturbed if \(\text{nn}(x)\) yields an anchor of the same type as \(x\), thus creating sparsity in some plots. Examples were selected at random.