Can Quantum Machine Learning Outperform Classical Methods? An Automated Benchmarking Framework for Smart Grid Security
Year: 2026
Type: dissertation
Abstract: Can quantum machine learning outperform classical methods for smart grid cybersecurity? On classical grid data, the debate remains unresolved. Proponents argue quantum methods leverage exponentially large entangled feature spaces to detect patterns in classical data inaccessible to classical approaches. Skeptics counter that quantum methods offer no advantage on classical data and that well-designed classical techniques remain superior. This thesis aims to resolve the debate through a principled evaluation framework that produces valuable outcomes regardless of which camp proves correct. The core contribution is an unbiased benchmarking framework comparing quantum and classical approaches. The Effective Quantum Proportion (EQP) metric quantifies which computational primitives survive evolutionary pressure based purely on performance. The Activation Function Categories (AFC) framework extends this by tracking subcategories across quantum and classical families. This delivers positive results irrespective of quantum's performance: if quantum methods excel, we identify genuine advantages; if they fail, evolutionary search automatically discovers superior classical alternatives. Applied to smart grid cyberattack detection across escalating realism, the framework consistently validates the skeptics' position. Evolutionary pressure systematically eliminated quantum operations in favor of classical alternatives. However, the search yielded a significant discovery: periodic inductive biases encoded in classical activation functions like Snake dramatically improve detection on power system time series. This demonstrates the framework's dual value—rigorous quantum evaluation coupled with automatic discovery of better classical methods. Given quantum machine learning's limitations on classical data, we examine scenarios where theoretical guarantees predict genuine advantages: quantum data from quantum sensors. Yet experiments with quantum-labeled periodic waveforms reveal substantial sensitivity to input structure, demonstrating that artificial quantum sensors cannot reliably predict real-sensor performance. Meaningful progress requires characterizing actual quantum measurement device outputs rather than synthetic benchmarks. This necessity for real quantum data motivates exploring hardware integration strategies. We present a preliminary hardware-first architecture demonstrating how quantum communication and sensing can be deployed where advantages are proven, while deferring quantum machine learning until realistic sensor data becomes available. This roadmap provides a foundation for future real hardware experiments and establishes methodology for rigorous quantum evaluation.
Source: TSpace (University of Toronto)
Author: Nader, Andrew
Institution: University of Toronto
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7203956749