Towards Practical Interrupt Side-Channel Attacks on macOS for Apple Silicon

In recent years, the high-end CPU landscape has shifted from an x86-dominated market to one increasingly featuring heavyweight Arm designs, most notably Apple's Mseries. While prior work has examined side-channel resilience in the Apple ecosystem, interrupt-based channels remain an open and unexplored area. This is due to Apple's proprietary hardware and software implementation, the practicality of interrupt side channel attacks are constrained by two open challenges: precise interrupt detection and having a clear understanding of the interrupt delivery mechanism. In this paper, we propose TIDE, a precise interrupt detection technique that exploits an explicit macOS behavior arising from Apple's Double Map mitigation. Specifically, Double Map overwrites the user-visible x18 register during kernel mapping restoration and interrupt handler branching on user-to-kernel transitions. To prevent kernel information from being accessed by user-space, macOS further clears x18 on kernel-to-user transitions in a subsequent update. Leveraging this behavior, TIDE enables two primitives without reliance on architectural timers: measuring intervals between consecutive interrupts and filtering interrupt-induced noise. Using TIDE, we reverse-engineer Apple's closed-source interrupt delivery mechanism and reveal that, unlike Linux, Apple's interrupt controller uniformly distributes shared peripheral interrupts across all active cores. We then demonstrate the benefits of both TIDE itself and the reverse-engineering results through three case studies. First, we fingerprint websites running on Safari with an accuracy of 93.8% in a closed-world setting and 91.2% in an open-world setting, and video fingerprinting with an accuracy of 78.1%. Next, we use TIDE to filter out interrupt noise in countingthread timers, where TIDE tracks the number of noisy interrupts and only introduces an overhead of only 0.06%. With this, we have sigificiantly improved the SysBumps attack by increasing its success rate from 54% to 81%. Third, we examine the effect of our reverse-engineering results on the loop-counting attack, and find that a compute-intensive thread that generates no interrupts can reduce its accuracy from 94.8% to only 39.3%. By augmenting the training dataset with the number of active cores, we successfully enhance the robustness of the attack, achieving accuracies to 92.8% and 93.6% with and without noise, respectively. Further, we apply TIDE to build a timer-less covert channel with a high bandwidth of 111.65 b/s. Finally, TIDE successfully extracts the keys from Cloudflare's Interoperable Reusable Cryptographic Library (CIRCL) v1.1.