Effective Universal Unrestricted Adversarial Attacks Using a MOE Approach

Combining Attack Success Rate and DetectionRate for effective Universal Adversarial Attacks

“…We propose a nested-evolutionary algorithm for generating universal unrestricted adversarial examples in a black-box scenario inspired by [13]. Given a sequence of image filters as input, the algorithm returns the best image-agnostic filter configuration which, applied to the images from the dataset, maximizes the classification error of the target model while the detection rate and the perturbation applied are minimized.…”

Lie to Me: Shield Your Emotions from Prying Software

“…The diffusion and the wide use of deep learning methods for artificial intelligence systems, thus, pose significant security and privacy issues. From the security point of view, Adversarial Attacks (AA) showed that deep learning models can be easily fooled [ 13 , 14 , 15 , 16 , 17 , 18 , 19 , 20 , 21 , 22 , 23 , 24 , 25 ] while, from a privacy point of view, it has been shown that information can be easily extracted from dataset and learned model [ 26 , 27 , 28 ]. It has also been shown that attacking methods based on adversarial samples can be used for privacy-preserving purposes [ 29 , 30 , 31 , 32 , 33 ]: in this case, data are intentionally modified to avoid unauthorized information extraction by fooling the unauthorized software.…”

“…The fooling protecting filters are built by composing and parametrizing popular image-enhancing Instagram filters: they are the result of an optimization process implemented by a nested-evolutionary algorithm [ 13 , 14 , 33 ]. Applying these protecting filters to any image, we obtain a series of other images from which information extraction is more difficult.…”

“…The proposed algorithm combines the idea of a Multi-Objective Evolutionary (MOE) approach for adversarial attacks [ 13 ] with the per-instance approach presented in [ 33 ]. Compared to the MOE approach, we also introduce the use of a Full-Reference Image Quality Assessment (FR-IQA).…”

“…As proposed in [ 13 , 14 , 33 ], we have used popular Instagram image filters such as Clarendon, Juno, Reyes, Gingham, Lark, Hudson, Slumber, Stinson, Rise, and Perpetua . Each filter has distinct properties and aspects, such as different contrast, saturation, brightness, and shadow levels.…”