Deep Learning Backdoor Attacks Detection

The susceptibility of deep neural networks to backdoor or trojan attacks has been demonstrated, wherein an adversary embeds a trigger during the training phase. This trigger allows the model to correctly classify regular inputs but produces a targeted and incorrect classification when the input contains the trigger. In this report, a trojan detection method was discussed, which circumvents the need for access to the training/test data, avoids computationally intensive operations, and does not rely on assumptions about the trojan trigger’s characteristics. Instead, this approach focuses on analyzing the weights of the network’s final linear layer. Empirical findings revealed several recurring traits in trojaned networks, absent in benign networks.