Deep-fool attack detection using dual-branch classifier

Deep neural networks have achieved remarkable success with regards to performance across all vision-driven applications. However, they still remain highly susceptible to adversarial attacks, which introduce imperceptible and visually indistinguishable perturbations capable of manipulating model predictions. The DeepFool attack, in particular, generates minimal perturbations by iteratively shifting the samples across decision boundaries, hence making its detection most challenging. In order to handle this vulnerability, this paper presents a dual-branch DeepFool detection framework and systematically evaluates the effectiveness of combining spatial features with MSCN-based statistical representations, which leverages both raw spatial features and natural image statistical cues based on MSCN coefficients. The proposed model has been experimented on the CIFAR-10 and MNIST datasets with DeepFool-generated examples, demonstrating that the performance significantly outperforms a baseline CNN in generalization and attack detection. A dual-branch classifier achieves up to 99.76% detection accuracy for deepfool samples and maintains accuracy close to clean data performance, whereas the MSCN branch exposes the perturbation artifacts, which are invisible in the original spatial domain. The results establish that integrating hand-crafted statistical features with deep representations offers an effective defense mechanism against the subtle deepfool attack, providing an empirically validated approach for improving detection of minimal perturbation attacks such as DeepFool.