Exécution de notre outil de synthèse sur un nouvel article

Résumé : Wireless networks have enhanced communication by making it seamless and accessible, however, they have introduced significant privacy vulnerabilities. Adversaries can passively monitor and analyze network traffic to identify devices and user activities. While traditional identifiers such as IP and MAC addresses can be easily spoofed, machine-learning approaches leverage traffic characteristics like packet sizes, timing, and protocol usage to fingerprint devices with high accuracy. This capability threatens user privacy by exposing sensitive behavioral patterns and personal habits. Although several countermeasures such as traffic padding, shaping, and cover traffic injection have been proposed to counteract such adversarial profiling their effectiveness remains limited, particularly against internal adversaries with full network visibility. In this study, we propose GANFUSION, a GAN-based traffic generation and injection framework designed to confuse machine learning classifiers. By emulating diverse device behaviors, GANFUSION masks genuine traffic patterns, making accurate identification more difficult for adversaries. Experimental results demonstrate that GANFUSION significantly reduces device identification accuracy and classifier confidence across all devices, outperforming existing countermeasures. It maintains 99.6% packet correctness while introducing only 83 bytes of bandwidth overhead, making it both effective and efficient for real-world privacy protection. Furthermore, it intelligently selects specific devices to increase uncertainty in classification outputs, enhancing resistance against network-based profiling attacks.