This seminar, presented by Gururaj Saileshwar, focuses on the critical need for robustly mitigating microarchitectural vulnerabilities like Spectre and Rowhammer by adopting a "good offense" (proactive testing and attack mindset) during the design phase.
The first part introduces AMuLeT (ASPLOS’25), an automated testing framework designed to detect speculative leakage vulnerabilities in microarchitectural defenses early in the design cycle, integrated with the Gem5 simulator. AMuLeT successfully uncovered six previously unknown vulnerabilities across four popular Spectre countermeasures, demonstrating its effectiveness in preventing the deployment of flawed defenses.
The second part addresses Rowhammer attacks, specifically in the context of DDR5’s Per-Row Activation Counting (PRAC) framework. Saileshwar details newly discovered "fill-and-escape" attacks that bypass existing PRAC implementations (like Panopticon) by exploiting the time gap between alert and mitigation. To counter this, QPRAC (HPCA’25) is proposed, which employs a priority-based mitigation service queue, co-designed with opportunistic and proactive mitigations, to ensure security at minimal performance overheads. The talk also highlights how PRAC's "Alert Backoff" (ABO) signals can inadvertently create new side and covert channels, allowing attackers to spy on victim activity and even leak secret keys.
Summary:
0:42 – The Arms Race in Microarchitectural Security: Gururaj Saileshwar emphasizes the critical need to robustly mitigate Spectre and Rowhammer vulnerabilities at the hardware level, as they undermine entire system security. He highlights the ongoing "arms race" between attackers (difficulty estimating threat landscape) and defenders (difficulty building robust, practical defenses without performance trade-offs), leading to defenses breaking down over time.
3:58 – Ending the Arms Race with Offensive Tools: The speaker argues that the only way to get ahead is by developing and applying good offensive tools at the design stage to detect vulnerabilities early, preventing the deployment of insecure solutions.
5:35 – AMuLeT: Design-Time Spectre Leakage Detection: Introduces AMuLeT, an automated framework to detect speculative leakage vulnerabilities in microarchitectural defenses within simulators. It extends model-based relational testing to microarchitectural simulators like Gem5.
13:56 – Overcoming Simulation Challenges: To address challenges like slow simulation speed and vast search space, AMuLeT uses optimizations: restricting microarchitectural traces to realistic attacker observables (cache/TLB), implementing a persistent testing harness to avoid simulator startup costs, and artificially increasing contention on microarchitectural structures to amplify vulnerability detection.
20:27 – AMuLeT's Impact: AMuLeT successfully conducted the first systematic test campaign on four popular Spectre defenses (published 2018-2024), finding vulnerabilities in all of them. It uncovered three previously known and six unknown leakages, including demonstrating that the recent SpecLFB defense is insecure, often detecting them in seconds to minutes.
23:41 – Types of Vulnerabilities Found by AMuLeT: These include unimplemented protections on certain execution paths, implementation bugs (e.g., incorrect L1 cache miss handling), undocumented performance optimizations (e.g., removing protections for first speculative load), and fundamental design issues like race conditions between speculative and non-speculative loads.
34:12 – Rowhammer: A Persistent Threat: The problem of Rowhammer (bit flips in neighboring DRAM rows due to rapid activation) persists, with activation thresholds decreasing significantly across technology nodes, necessitating scalable solutions.
35:24 – Per-Row Activation Counting (PRAC): JEDEC's DDR5 standard introduced PRAC, a principal mitigation technique. It uses per-row counters, alerts the CPU when a row crosses a threshold, and the CPU sends an RFM command for the DRAM to mitigate neighboring rows.
37:25 – The "Tabbo_act" Gap and Vulnerabilities: A critical time gap (Tabbo_act, 180ns) between the DRAM detecting an attack and the CPU sending mitigation allows attackers to perform additional activations. Prior PRAC implementations (like Panopticon's FIFO queue) fail to properly track pending mitigations, leaving them vulnerable.
41:25 – Fill-and-Escape Attack: A new attack against Panopticon is demonstrated, exploiting the Tabbo_act gap and FIFO queue. Attackers fill the queue with "dummy" rows, then hammer a "target" row when the queue is full, preventing it from being added for mitigation and allowing its activation count to surpass the threshold.
45:28 – QPRAC: A Robust Rowhammer Defense: QPRAC (HPCA’25) proposes a more robust PRAC implementation using a priority-based mitigation queue that tracks rows by their activation counts. This ensures that the most heavily activated rows are always prioritized for mitigation.
46:54 – QPRAC Performance: QPRAC, combined with opportunistic (all banks mitigate on alert) and proactive (mitigating on TF) strategies, achieves near-negligible performance slowdowns (less than 1%) compared to a PRAC-enabled baseline.
1:01:10 – New Attacks from PRAC Defenses: Even new mitigations can introduce new vulnerabilities. PRAC's "Alert Backoff" (ABO) signals, which freeze an entire DRAM rank, are visible across the system and can be exploited as powerful microarchitectural side channels.
1:03:59 – Spying and Information Leakage: Attackers can observe latency spikes caused by victim activity triggering ABOS, allowing them to spy on victim memory access patterns and even quantify the exact number of activations a victim performed on a given row.
1:07:20 – Side and Covert Channels: This information can be weaponized for side-channel attacks (e.g., an AES table attack leaking secret key bytes) and covert channels (transmitting 10-100 kilobits per second) by encoding information in activation counts or activity.
1:08:54 – Defending Against New Side Channels: Proposed defenses include replacing activity-based mitigations with timing-independent ones (fixed or variable rate RFMs) and co-designing them with proactive mitigations on TF, aiming for low overheads.
1:10:06 – Conclusion: Proactive Security is Key: The speaker concludes by emphasizing that security is crucial for next-generation systems, and an "offensive mindset" applied at design time is essential for creating truly robust and long-lasting defenses.