Ramesh Babu Battula

In this paper, we introduce a new, lightweight, and secure solution, PrivEdge-Loc, which is particularly suitable for multi-band indoor localization in IoT networks. Unlike traditional federated learning, which requires substantial computational power in edge computing, Split Learning greatly reduces this necessity in our solution. Split Learning is a novel machine learning paradigm in which the neural network is split into two parts: one on the edge device and the other. By offloading heavy computation to the server, the edge nodes only need to send low-dimensional compressed data to the server for processing, ensuring that the privacy of raw Channel State Information (CSI) is architecturally protected against eavesdropping. To tackle the statistical heterogeneity in multi-frequency environments, the server leverages a differentiable gating network within a dynamic Mixture of Experts (MoE) architecture. This enables the server to dynamically route incoming features to frequency-specific experts without requiring explicit band identification. Additionally, the representational power of the student experts is further enhanced through Hint-Based Knowledge Distillation (HKD), which actively aligns the student experts’ intermediate latent space with that of the pretrained teacher models. Our proposed PrivEdge-Loc is evaluated using a non-IID dataset with 3 frequency bands: 2.4 GHz, 3.5 GHz, and 5.0 GHz. Our framework achieves a highly competitive Mean Absolute Error (MAE) of 0.194 meters. The proposed framework provides a highly scalable, secure, and feasible alternative to computation-intensive federated baselines for data-sensitive, multi-tenant deployments.

Split Learning (SL) is a distributed learning framework that offers lightweight collaborative deep learning on resource-constrained devices. However, SL's distributed nature makes it vulnerable to client-side backdoor attacks. In this scenario, an adversary client can maliciously influence the global model on the server. The existing defenses heavily rely on the spatial consensus mechanism, which compares client updates and filters out statistical outliers. We demonstrate that such defenses do not hold in a non-IID data distribution environment, where honest clients naturally have gradients with high divergence. Thus, the state-of-the-art defenses suffer from a high false positive rate during the training rounds. This results in excluding benign participants and causing model instability at convergence. To this end, the proposed GradSent replaces spatial agreement with temporal consistency. The key idea is that although benign clients may differ from one another at each step, they tend to follow a stable and convergent trajectory when viewed across the training history. GradSent captures this behavior through two phase framework that first learns the manifold of benign dynamics with a lightweight Gradient Autoencoder (GAE) and then evaluates new updates against this learned temporal profile. It further employs iterative self-cleaning to remove possible contaminations from the burn-in history and dynamic thresholding to adapt to shrinking gradients. Our experiments on various datasets and architectures demonstrate that GradSent effectively blocks most of the client-side backdoor attacks while preserving Main Task Accuracy under extreme non-IID data distributions.

The swift growth in ubiquitous connectivity, such as 5G and 6G, cloud computing, which enables data-driven automation and increases efficiency across industries, gave rise to the Internet of Things. This proliferation significantly expands the network attack surface, making IoT-based networks increasingly vulnerable to a multitude of attacks. To mitigate the attacks, Intrusion detection systems are developed. While developing the IDS, a primary challenge is the data collection of attack patterns and their training. The data collected is often massive, heterogeneous, and, critically, imbalanced. Training conventional models on such imbalanced data typically leads to poor performance, particularly in overlooking minority attack classes during multi-class classification. Our proposed system directly addresses this challenge of data imbalance. We introduce a hybrid technique, the Transformed Image Generative Adversarial Network (TIGAN), that generates high-fidelity attack samples and is then used to augment underrepresented attack classes, complemented by downsampling of overrepresented ones. The balanced data allows our lightweight deep learning model, Transformed Image Convolutional Neural Network (TICNN), to achieve robust attack detection. The experimental TICNN-based IDS delivers a faster detection time and detection accuracy of 99.56% at the macro level, which showcases its strong capability to identify diverse attacks, sporadic minority classes, in challenging IoT settings.