From Assault to Protection: Insights into Deep Studying Safety Measures in Black-Field Settings
Authors: Firuz Juraev, Mohammed Abuhamad, Eric Chan-Tin, George K. Thiruvathukal, Tamer Abuhmed
Summary: Deep Studying (DL) is quickly maturing to the purpose that it may be utilized in safety- and security-crucial functions. Nonetheless, adversarial samples, that are undetectable to the human eye, pose a severe risk that may trigger the mannequin to misbehave and compromise the efficiency of such functions. Addressing the robustness of DL fashions has grow to be essential to understanding and defending towards adversarial assaults. On this examine, we carry out complete experiments to look at the impact of adversarial assaults and defenses on numerous mannequin architectures throughout well-known datasets. Our analysis focuses on black-box assaults equivalent to SimBA, HopSkipJump, MGAAttack, and boundary assaults, in addition to preprocessor-based defensive mechanisms, together with bits squeezing, median smoothing, and JPEG filter. Experimenting with numerous fashions, our outcomes show that the extent of noise wanted for the assault will increase because the variety of layers will increase. Furthermore, the assault success price decreases because the variety of layers will increase. This means that mannequin complexity and robustness have a big relationship. Investigating the variety and robustness relationship, our experiments with numerous fashions present that having a lot of parameters doesn’t indicate larger robustness. Our experiments lengthen to point out the consequences of the coaching dataset on mannequin robustness. Utilizing numerous datasets equivalent to ImageNet-1000, CIFAR-100, and CIFAR-10 are used to guage the black-box assaults. Contemplating the a number of dimensions of our evaluation, e.g., mannequin complexity and coaching dataset, we examined the conduct of black-box assaults when fashions apply defenses. Our outcomes present that making use of protection methods can considerably scale back assault effectiveness. This analysis offers in-depth evaluation and perception into the robustness of DL fashions towards numerous assaults, and protection
