Content area
The relentless pursuit of performance has driven significant innovation in computer architecture, yielding systems that are faster, more efficient, and increasingly specialized. This momentum is fueled by the growing heterogeneity of modern applications, which demand tailored, domain-specific optimizations. However, this performance race comes at a cost, as many optimizations inadvertently introduce covert channels - unintended paths that can inadvertently leak sensitive information. As more computation migrates to the cloud, this attack surface expands, raising a fundamental question: “How secure is our information?” Alarmingly, even when data is encrypted, adversaries can often infer private details by observing system behavior. Over the past decade, a wave of research has revealed how subtle interactions between software and hardware can give rise to side-channel vulnerabilities, enabling attackers to extract secrets without breaking cryptographic guarantees. These developments have made it increasingly difficult to simultaneously achieve high performance and strong security. This work argues that to achieve both trustworthiness and efficiency, system architects must rigorously account for the trade-offs between performance and security.
To support this argument, this dissertation investigates three scenarios where this trade-off is pronounced. First, it addresses the challenge of sharing program traces for performance analysis without compromising trace security. Current techniques for generating realistic traces capture a range of behaviors necessary to be evaluated, containing a lot of information about the application, its inputs and the underlying system on which it was generated. Consequently, generating traces from real-world executions risk leakage of sensitive information. To prevent this, traces can be obfuscated before release. However, this can undermine their ideal utility, i.e., how realistically a program behavior was captured. To bridge this gap, we develop Camouflage, a trace obfuscation framework that systematically balances trace utility and input confidentiality through semantically guided transformations. Second, it examines Fully Homomorphic Encryption - first through the lens of privacy, and then from a performance characterization perspective. It reveals that encrypted computations remain vulnerable to side-channel leakage through memory access patterns. Separately, it presents CryptOracle, a modular framework that enables fast, interpretable performance estimation of FHE workloads without the need for full-system simulation. Finally, the dissertation explores vulnerability of on-chip interconnects, where existing defenses designed for core and cache interaction are rendered inadequate. It shows that while these networks are optimized for resource sharing, their contention patterns can be manipulated to form microarchitectural side-channels. A temporal isolation defense is adopted to enforce non-interference of secure and adversarial traffic. Together, these contributions advocate for a new design philosophy : one that treats performance and security not as competing objectives, but as deeply intertwined aspects of modern system design. By systematically characterizing these trade-offs and introducing frameworks to navigate them, this dissertation lays the foundation for designing future systems that are both high-performing and secure by design.