سروش هاشمي فر

With the development of Deep Neural Networks (DNNs), their need for computing resources increases, and this has increased the popularity of outsourcing the network training. However, training on third-party platforms may expose the network to potential risks that an attacker could deliver during a adversarial attack on an infected network with a backdoor that behaves normally when faced with clean samples; But adversarial samples can trigger this backdoor and trick the model into classifying it into the attacker's target label when the trigger pattern appears in the network's inference process. In this regard, the previous works have limitations such as the need for a huge defense dataset and their dependence on the characteristics of the Trojan trigger. Their performance deteriorates in the face of a limited defense dataset. This thesis propose an integrated method to detect the backdoored model and clean it with limited access to the defense data, ignorance of the network learning process and the characteristics of the Trojan trigger. Based on the sensitivity of the neuron to the input, its contribution on network predictions, its activity and the correlation of this activity with other neurons, the probability of contamination of each neuron is determined by the proposed method. Then the backdoored neurons are pruned from the network. Due to the examination of the characteristics of each neuron, the output of the proposed method can be interpreted and the pruning of the neurons can be justified. Also, since the proposed detection and removal method is integrated, both can be implemented simultaneously in one process. The eva‎luations carried out using the CIFAR-10 dataset on the ResNet-18 network architecture and against four different attacks show that the proposed method cleans 97.6% of the backdoor footprint from the network on average. On the other hand, this amount of contamination removal is obtained with 4% degradation in model accuracy and 0.39% pruning of network neurons on average. In addition, the proposed method performs more than three times faster than the advanced data-driven methods regarding execution time.

دفاع در برابر بدافزار عصبي , پاك‌سازي درب پشتي , حملات خصمانه , شبكه‌هاي عصبي پيچشي , هوش مصنوعي تفسير‌پذير

Neural Trojan Defense , Backdoor Mitigation , Adversarial attacks , Convolutional Neural Networks , Explainable AI

Soroosh Hashemifar

Dr. saeed Parsa