Omni: automated ensemble with unexpected models against adversarial evasion attack

Simulation Driven Intelligent Business Data Analysis and Decision-Making

“…Intelligent decision-making technologies that introduce artificial intelligence, machine learning, and other methods have emerged. How to assist enterprises in responding to development changes, improving operational levels and benefits, and promoting highquality business development and growth, becoming a key topic of general concern for enterprises and society at present [11].…”

Adversarial attacks on YOLACT instance segmentation

“…In the context of security applications, Shu et al [26] employed the Omni method to create unexpected sets of models whose hyperparameters are controlled to make the attacker's target model far away, for defensive purposes. Wang et al [54] proposed Def-IDS, an ensemble defense mechanism capable of resisting known and unknown adversarial attacks, composed of two parts: a multi-class generative adversarial network and a multi-source adversarial retraining technique.…”

“…The controversy leads to the first question: Are general model ensembles and ensembles defenses guaranteed to be more robust than individuals? Existing ensemble evaluation methods [26], [27] are insufficient to answer this question primarily due to three aspects: (i) they are largely gradient-based [11], [26], [27] without systematically studying gradient-free methods, which are more suitable for cybersecurity discrete data. (ii) they focus on homogeneous ensembles that use deep [26], [27] and tree models [28], but neglecte heterogeneous ensembles that combine deep and tree models.…”

“…Existing ensemble evaluation methods [26], [27] are insufficient to answer this question primarily due to three aspects: (i) they are largely gradient-based [11], [26], [27] without systematically studying gradient-free methods, which are more suitable for cybersecurity discrete data. (ii) they focus on homogeneous ensembles that use deep [26], [27] and tree models [28], but neglecte heterogeneous ensembles that combine deep and tree models. (iii) they explore the ensemble of multiple attack methods [22], [27], [29] more, and rarely examine the ensemble of perturbations from other models (transfer ensemble attack) [22].…”

Fast and Robust Detection of Adversarial Attacks in the Problem Space using Machine Learning

“…[41] Without learning how organizers constructed the adversarial cases, I used the defense mechanism to achieve the AICS'2019 Challenge. [42] Using techniques like hyperparameter optimization, it is recommended to generate an ensemble featuring profound models different from the attacker's expected framework (i.e., targeting model). [43] proved that the unique adversarial training greatly improves the robustness of deep learning models against a large variety of attacks.…”