SECD: Detecting Adversarial Examples with Multi-Strategy AI Techniques
Автор: AI-WEINBERG
Загружено: 2026-01-06
Просмотров: 3
Описание:
Similarity Ensemble Contradiction Detection (SECD): A Multi-Strategy Framework for Adversarial Example Detection
By Abraham Itzhak Weinberg
Deep neural networks are vulnerable to adversarial examples—tiny, carefully crafted changes to inputs that cause models to make mistakes. We introduce Similarity Ensemble Contradiction Detection (SECD), a new framework that detects adversarial examples by spotting contradictions between predicted classes and similarity-based evidence from the training data.
SECD uses three main strategies:
1. An ensemble of complementary similarity metrics in both feature and pixel space.
2. Class centroid analysis to detect unusual patterns in feature space.
3. Integration with feature squeezing for stronger, more robust detection.
On MNIST, SECD achieves a 54% detection rate with only 8.8% false positives. On CIFAR-10, detection is more challenging due to natural image variability, reaching 20.9% detection at 9.6% false positives. These results provide insights into the limits of similarity-based adversarial detection and highlight how performance depends on dataset complexity.
Our work shows that while SECD is effective for well-separated feature spaces, natural images still pose fundamental challenges for adversarial detection.
Повторяем попытку...
Доступные форматы для скачивания:
Скачать видео
-
Информация по загрузке:
