Forough Shirin Abkenar

Forough Shirin Abkenar

Ph.D., Electrical and Information Engineering

WiFi Pathologies Detection using LLMs

Abstract

The encoder-only and decoder-only large language models (LLMs) are fine-tuned to detect pathologies in IEEE 802.11 networks, commonly known as WiFi. The approach involves manually crafting prompts followed by fine-tuning. Evaluations show that the sequential model achieves high detection accuracy using labeled data, while the causal model performs equally well for unlabeled data.

Motivations

  • The methods in the literature tailored to individual users, necessitating adjustments based on each user’s specific needs. However, the pathologies are common for all users, suggesting the potential for developing generalized approaches or models that can effectively address these shared issues without the need for extensive customization.

  • The above-mentioned methods need to access comprehensive information of the users’ devices that rises privacy concerns. Indeed, the critical information, such as detailed device data and network configurations, can be divulged, especially if such information is not essential for resolving WiFi issues. Thus, the existing research work are vulnerable to privacy violation of users information.

Contributions

  • The objective is to leverage LLMs for detecting pathologies in WiFi connections.

  • Given network parameters, we first apply prompt engineering to craft effec- tive prompts for pathology detection.

  • To demonstrate LLMs’ efficacy in network applications, an encoder-only model is fine-tuned for supervised classification of pathologies.

  • To minimize computational complexityemploy, parameter-efficient fine-tuning (PEFT) is employed whereby the low-rank adaptation (LoRA) is applied.

  • The approach is extended by fine-tuning a decoder model for unsupervised data, ensuring robust detection across different data scenarios.

Results

Energy
Delay
Delay

Performance of fine-tuned DistilBert model, LoRA-based model, and fine-tuned GPT2 model.

Conclusion

The effectiveness of LLMs was demonstrated in wireless communications for detecting WiFi pathologies. The encoder-only and decoder-only LLMs fine-tuned for supervised and unsupervised data, respectively, using manually crafted prompts tailored for pathology detection. Due to limited available data, the study focused on noise-driven pathologies, though existing research highlights additional causes, i.e., signal strength, hidden terminals, and capture effects.