Adversarial Robustness of Forward-Forward Models

Deep learning architecture has been demonstrated to be vulnerable in adversarial settings, where fine-tuned modifications to input instances, appearing benign to humans, can lead to model missclassifications. This characteristic has been extensively studied in computer vision, where no defense to date is able to fully mitigate these attacks. This weakness appears to be inherent to models’ architecture and a direct consequence of their training process, which relies on gradient descent combined with backpropagation. This limitation is the result of the differences between brain and model function leading to unexpected decision boundaries. To address this disparity, alternative perceptronbased architectures have been proposed, such as the forwardforward neural network that aims to be closer to brain behavior. This architecture has demonstrated high performance in noise resistance, raising questions about its robustness in adversarial settings and its potential to withstand attacks. The aim of this paper is to study the robustness of forward-forward models against adversarial attacks and show that, despite their noise resistance, they are also vulnerable in adversarial settings. To demonstrate this, we use optimization-based methods to craft adversarial examples capable of bypassing forward-forward models in $100 \%$ of the tested cases, with modifications comparable to those used on backpropagation-based neural networks. We explore the transferability properties between backpropagation and forward-forward neural network networks, showing that, even considering some disparity in terms of representation, these models share some common characteristics.