From Side-Channel Profiling to Adaptive Information Control A Learning-Driven Framework for Intelligent Side-Channel Defense

Side-channel analysis has traditionally been investigated as a mechanism for extracting sensitive information from physical leakages such as power consumption and electromagnetic emissions. While most existing research focuses on cryptographic key recovery, the growing deployment of embedded and Internet of Things (IoT) systems has expanded the relevance of side-channel techniques toward behavioral profiling, runtime monitoring, and security assessment. At the same time, increasing environmental variability, noise, and system complexity expose the limitations of static machine learning pipelines and fixed defensive strategies. This thesis presents a learning-driven framework for application-level behavioral inference based on side-channel observations, combining signal processing, machine learning, deep learning, and adaptive control mechanisms. Using electromagnetic emissions and current consumption measurements acquired from a Raspberry Pi-based platform, a complete processing pipeline is developed, including temporal segmentation, RMS-based representation, robust normalization, feature extraction, feature selection, and classification. Experimental evaluations conducted across multiple application scenarios demonstrate that side-channel information can be effectively exploited to identify runtime activities without access to software internals, network payloads, or cryptographic material. A systematic analysis of robustness under noisy conditions reveals that classification performance is strongly influenced by the stability of the feature representation. The results show that temporal aggregation through majority voting significantly improves session-level reliability, while feature relevance and class separability evolve dynamically as noise conditions change. These observations motivate a shift from static inference toward adaptive representation control. To address this challenge, the thesis introduces a feature-aware Reinforcement Learning framework operating directly within the feature space. Rather than acting as a classifier, the Reinforcement Learning agent functions as an adaptive controller that modifies feature representations according to observed instability, uncertainty, and classifier feedback. Experimental results demonstrate that adaptive feature-space control can improve robustness under noisy conditions while preserving interpretability and deployment flexibility. Beyond its practical contributions, this work proposes a conceptual transition from side-channel profiling to adaptive informational defense. The same adaptive mechanisms used to stabilize legitimate inference can also be employed to reduce the exploitable informational content available to adversarial observers. Building upon this observation, the thesis introduces the foundations of Adaptive Inferential Security (SIA), a novel security paradigm in which protection is achieved through the dynamic control of inferential observability rather than through access restriction alone. Within this perspective, security becomes a problem of regulating what can be inferred from observable system behavior, creating an intentional asymmetry between legitimate and adversarial inference processes. The results demonstrate that side-channel analysis can evolve beyond passive observation and static classification toward adaptive, learning-driven control of information leakage. In doing so, this thesis establishes an experimental and conceptual foundation for future intelligent security systems capable of dynamically managing their own inferential exposure in embedded, cyber-physical, and IoT environments.