Adversarial Vulnerabilities and Defenses in Social Emotion Recognition

In this paper, we explore the robustness of a CNN model that is trained using the AffectNet dataset. We use convolutional neural network (CNN) as a base to examine how adversarial perturbations can impact the performance of the model, particularly focusing on two widely used attack strategies: First one is ’Fast gradient sign method’ and the second one is ’Projected gradient descent’. These methods introduce small but specific changes to the input images, which result in changes in the output class and, in consequence, degrade the performance of the model. To address these challenges, we apply two defense mechanisms aimed at mitigating the effects of these attacks: adversarial training perturbs the model’s training process with adversarial examples, and defensive distillation involves training the model to produce more smooth class probabilities. In verifying the efficacy of these defense approaches, we illustrate how they can help in reinstating the model’s accuracy as well as shield against adversarial threats. Specifically, it is not the CNN architecture that is the focus of this paper but on how defense strategies work to protect what is often the main concern when working with such networks: reliable emotion recognition in the face of adversarial forces.