Pixel Power: Harnessing Image Processing for Optimizing Few-Shot Malicious Traffic Detection
Yihan Wang, Ruiyi Yan · 2024
The rapid proliferation of cyber threats in an increasingly digitized world presents a formidable challenge to existing cybersecurity mechanisms of which malicious traffic is a noticeable issue. However, feature extraction of malicious traffic is often hindered by insufficient data, and thus training effectiveness and efficiency are compromised. To address the challenge of detecting malicious traffic with small sample sizes, we introduce a meta regularization network, MRN. Our method leverages meta-learning techniques alongside convolutional neural networks to capture intricate traffic patterns, so as to classify processed encrypted traffic data. Within MRN, we introduce a novel meta-learning regularization strategy that adjusts the training process’s loss based on network parameters, in order to improve computing efficiency. Our comprehensive experimental evaluations reveal the model’s exceptional capability in malicious traffic classification, achieving 95.11% accuracy averagely, which is higher than the baseline methods under various experimental setup. This underscores our framework’s potential in advancing the field of cybersecurity through image-based analysis and small-sample classification techniques.