Data Privacy Analysis for ML-Based Congestion Control in Multi-Hop Wireless Networks

Control protocols play a pivotal role in ensuring the efficient operation of communication networks, providing services across various layers of the network architecture. Modern protocols are increasingly incorporating Machine Learning based algorithms, a trend that has brought data privacy concerns to the forefront. This paper addresses the challenge of safeguarding the privacy of the input data utilized by ML algorithms within network control protocols, particularly focusing on a specific congestion control protocol for multi-hop wireless networks. Within this protocol, the values disseminated by each node can inadvertently disclose sensitive information about the equipment or the network state. We explore techniques to mitigate this privacy risk while acknowledging the trade-off between enhanced privacy and data utility. The study examines the ensuing performance implications, shedding light on the balance between privacy protection and algorithm accuracy. We propose a stochastic framework and conduct an extended set of simulations to compare four different privacy-preserving techniques. The added privacy is analyzed using Kullback-Leibler divergence and cosine similarity, while the impact on protocol performance is evaluated through Root Mean Square Error and Jain Fairness Index. The results provide a range of values that allows network operators to select the optimal configuration based on their specific privacy requirements, demonstrating that adequate levels of privacy and performance can be achieved simultaneously within this framework.