AIにより推定されたラベル
※ こちらのラベルはAIによって自動的に追加されました。そのため、正確でないことがあります。
詳細は文献データベースについてをご覧ください。
Abstract
Making classifiers robust to adversarial examples is hard. Thus, many defenses tackle the seemingly easier task of detecting perturbed inputs. We show a barrier towards this goal. We prove a general hardness reduction between detection and classification of adversarial examples: given a robust detector for attacks at distance (in some metric), we can build a similarly robust (but inefficient) classifier for attacks at distance /2. Our reduction is computationally inefficient, and thus cannot be used to build practical classifiers. Instead, it is a useful sanity check to test whether empirical detection results imply something much stronger than the authors presumably anticipated. To illustrate, we revisit 13 detector defenses. For 11/13 cases, we show that the claimed detection results would imply an inefficient classifier with robustness far beyond the state-of-the-art.