Introduction
As the semiconductor industry shifts toward 2.5D chiplet-based designs to overcome the scaling limitations of monolithic chips, a new security frontier is emerging. While modularity improves yields and lowers costs, it introduces a significant supply chain risk: the integration of third-party chiplets that may harbor Hardware Trojans (HTs). This research explores how these malicious components can exploit cache coherence protocols to silently manipulate memory, bypassing traditional software-level security and physical location constraints.
The Vulnerability: Permissionless Coherence
Modern computing systems rely on cache coherence protocols to ensure that all processing cores see the most up-to-date version of data. However, these protocols are primarily designed for performance and correctness, not security. They typically operate "below" the operating system and do not enforce memory permissions, such as virtual versus physical address space restrictions.
Previous research on Hardware Trojans often focused on Network-on-Chip (NoC) attacks, where a Trojan modifies data packets in transit. Such attacks are limited by physical proximity; the Trojan must be located on the direct path between the victim core and the memory controller. In contrast, if a Trojan can "speak" the coherence protocol, it can manipulate any memory region in the system regardless of its physical location on the interposer. As long as the malicious chiplet is connected to the coherence fabric, it can broadcast requests that the system's global directory treats as legitimate.
Methodology and Attack Taxonomy
The study categorizes and simulates several attack vectors targeting the MOESI Hammer protocol—a hybrid broadcast-directory protocol common in modern high-performance processors. The researchers defined a taxonomy of four basic architectural attacks:
- Passive Reading: Snooping on coherence messages to leak sensitive data.
- Masquerading: Spoofing the sender ID of a packet to cause system deadlocks or data incoherence.
- Modification: Altering message types (e.g., upgrading a shared request to an exclusive one) to disrupt state machines.
- Diverting: Blocking a legitimate request and rerouting it to a malicious destination.
The Forging Attack
The core contribution of this work is the "Forging Attack," a sophisticated two-phase orchestration designed to corrupt memory silently:
- Phase 1 (Acquiring Access): The Trojan snoops the coherence traffic for a specific target memory address. Once identified, it generates a malicious "Get Exclusive" (
GETX) request. The system's global directory, seeing a protocol-compliant request, invalidates the victim's copy and grants ownership of the cache line to the Trojan's chiplet. - Phase 2 (Writing Malicious Data): The Trojan performs a standard "write-back" (
PUTX) operation. It sends corrupted data to the memory controller, which writes the value to main memory. Because the Trojan followed the protocol, the system views this as a standard memory update.
Experimental Setup
The researchers evaluated these threats using the gem5 architectural simulator. The simulation environment modeled a 64-core system based on the Rocket-64 architecture, distributed across eight chiplets in an interposer-based 2.5D layout. To test the attack, they utilized a custom victim application that iterates over an array, providing a predictable target for the Trojan to corrupt.
Key Results and Insights
The simulation successfully demonstrated the viability of the Forging Attack. The Trojan intercepted a memory line owned by the victim process and modified a specific value (changing a '1' to a '5').
Several critical insights emerged from these results:
- Transparency: The attack is completely transparent to the software layer. No OS-level security triggers or exceptions were raised because the hardware was technically behaving according to the coherence protocol.
- Protocol Compliance: All malicious messages were "legal" within the MOESI Hammer framework. Standard hardware checkers designed to catch protocol errors would not flag these transactions.
- Location Independence: Unlike NoC-based Trojans, the malicious chiplet did not need to be "in the middle" of the victim's data path. It successfully initiated the attack by simply being a participant in the global coherence fabric.
Conclusion and Future Work
The research highlights a fundamental gap in hardware security: the lack of authentication in coherence traffic. While the Forging Attack was tailored to the MOESI Hammer protocol, the underlying principle—that coherence is permissionless—applies to most modern directory-based systems.
Future defensive strategies may involve hardening the 2.5D interposer. Since all chiplet-to-chiplet communication passes through the interposer, it could serve as a "Hardware Root of Trust." By implementing runtime monitoring within the interposer, systems could observe coherence packets for suspicious patterns and identify malicious behavior originating from untrusted chiplets before data corruption occurs.
Summary updated