TL;DR
As the semiconductor industry shifts from monolithic chips to "chiplet" designs connected via an interposer, security risks from untrusted third-party components are rising. This paper introduces GETXspy, a high-bandwidth covert-channel attack that exploits standard cache coherence protocols, and proposes a defense: an active interposer acting as a physical root of trust that filters and validates all inter-chiplet communication with only a ~4% performance overhead.
What Problem Does This Paper Address?
Modern processors are increasingly built using 2.5D integration, where various functional units (cores, accelerators, memory controllers) are manufactured as separate "chiplets" and then bonded to a silicon carrier called an interposer. While this improves yields and lowers costs, it creates a massive security gap: a single malicious or "buggy" chiplet sourced from a third-party vendor can potentially compromise the entire system.
The authors identify the cache coherence protocol—the hardware mechanism that ensures all cores see the same data in memory—as a primary attack vector. Because coherence is managed entirely by hardware and is largely invisible to software, traditional security measures (like OS-level permissions) often fail to see malicious interactions. The authors specifically target the lack of a "root of trust" in these distributed systems, where no single component currently has the authority to verify if a chiplet’s request for data is legitimate or part of an attack.
Methods
The paper approaches the problem from two sides: the attack and the defense.
The Attack: GETXspy
The authors developed GETXspy, a new covert-channel attack. In this scenario, a "spy" process on one chiplet generates specific write-ownership requests (GETX messages). Even if these requests are "legal" according to memory permissions, a hardware Trojan in a different chiplet can snoop on the resulting broadcasted invalidation messages. By modulating the addresses of these requests, the spy can exfiltrate data to the Trojan at a rate of 4.22 Mbps, bypassing existing security boundaries.
The Defense: Active Interposer & CMCs
To counter this, the authors propose using an active interposer (one containing active logic like routers) as a Root of Trust. Since all communication between chiplets must pass through the interposer, it is the perfect place to enforce security.
They introduce Coherence Message Checkers (CMCs) embedded within the interposer’s Network-on-Chip (NoC) routers. These CMCs consist of:
- Address Protection Unit (APU) Table: A hardware lookup table that stores memory access permissions (Read/Write/None) for every chiplet across different memory regions.
- Packet Checker/Modifier (PCM): Logic that intercepts every coherence packet. It verifies the requester's ID, the message type, and the target address against the APU table. If a chiplet tries to "snoop" on a region it doesn't own, the CMC-2 can convert a broadcast message into a directed "null" response, effectively blinding the attacker.
Datasets
The authors evaluated their design using a full-system simulation environment:
- Simulator:
gem5, a standard tool for computer architecture research. - Workloads: A subset of the SPEC 2006 benchmark suite (including benchmarks like
povray,hmmer, andsphinx3) to measure single-threaded and multi-programmed performance. - System Configuration: A 64-core system consisting of 8 chiplets (8 RISC-V cores each) using the MOESI Hammer coherence protocol.
Key Insights and Ideas
- Coherence is the "Soft Underbelly": Most security research focuses on the NoC or the OS, but the coherence protocol’s predictability and prevalence make it an ideal, overlooked target for hardware Trojans.
- The Interposer as a Chokepoint: By moving security logic into the interposer, the authors create a "secure-by-construction" environment. You don't need to trust the chiplets if you can trust the "floor" they stand on.
- Protocol-Agnostic Defense: The CMC approach doesn't require redesigning the complex coherence protocol itself. It simply polices the messages the protocol generates, making it easier to verify and implement.
- Filtering Improves Performance: Interestingly, the defense can actually reduce network congestion. By filtering out broadcast messages to chiplets that don't have permission to see them, the system saves bandwidth.
Key Results
- Attack Potency: The GETXspy attack achieved a bandwidth of 4.22 Mbps, which is more than enough to leak sensitive cryptographic keys or user data in real-time.
- Security Coverage: The proposed active interposer successfully blocked GETXspy as well as other common attack types like masquerading (spoofing a chiplet ID), modifying messages, and diverting data to unauthorized cores.
- Low Overhead: The average performance impact was only ~4%. In some multi-programmed workloads, the overhead actually decreased as the system scaled, because the CMCs' filtering of unnecessary broadcast traffic helped offset the latency of the checks.
Limitations and Future Work
- OS Dependency: The system relies on a "secure OS" running on a trusted co-processor within the interposer to manage the APU table. If the secure OS is compromised, the hardware defense can be reconfigured by an attacker.
- Intra-Chiplet Attacks: The defense only monitors traffic between chiplets. If an attack happens between two cores on the same physical chiplet, the interposer never sees the traffic and cannot stop it.
- Physical Attacks: The paper does not address physical side-channels (like power analysis) or DRAM-level attacks like Rowhammer, which would require additional, complementary defenses.
- Future Work: The authors suggest extending this model to heterogeneous systems (mixing CPUs with GPUs and FPGAs) and exploring more granular memory protection schemes.
Summary updated