Exploiting Shared Power Infrastructure in Multi-Tenant FPGAs

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 215 papers22 conceptsfrozen 2026-10-04

Cloud computing has traditionally relied on Field-Programmable Gate Arrays (FPGAs) as single-tenant resources. A single user rents a whole chip, ensuring that their logic and data are physically isolated from others. However, to maximize efficiency and reduce costs, the industry is moving toward multi-tenancy, where multiple independent users share the same physical FPGA chip.

While logical isolation (separating the "wires" of different users) is relatively straightforward, researchers have discovered a fundamental physical vulnerability: the Power Distribution Network (PDN). Because all users on a single chip draw from the same power supply, the electrical activity of one tenant can directly interfere with the operational integrity and data privacy of another.

The Shared Power Distribution Network

In a multi-tenant FPGA, users are assigned specific regions of the chip's fabric. While the routing of signals is strictly partitioned, the PDN is a global resource. Think of it like an apartment building where every unit shares the same water main. If one resident opens every tap at once, the water pressure drops for everyone else.

In the context of FPGAs, this "pressure" is voltage. As Muhammed Kawser Ahmed and colleagues explain in Multi-Tenant Cloud FPGA: A Survey on Security, this shared infrastructure creates a "hardware backdoor." A malicious tenant can use the PDN as a side-channel to observe or influence the behavior of their neighbors.

Power Hammering and Denial of Service

The most aggressive form of interference is a "power-hammering" attack. By deploying circuits designed to toggle at extremely high frequencies, an attacker can draw massive amounts of current.

In Denial-of-Service on FPGA-based Cloud Infrastructures - Attack and Defense, Tuan La and Dirk Koch demonstrate that these attacks can bypass existing security "fences." By drawing excessive current, an attacker can trigger the FPGA's internal safety mechanisms, causing the entire hardware instance to crash or go offline for extended periods. This effectively turns a shared resource into a weapon for Denial-of-Service (DoS).

Voltage Hammering and Fault Injection

Beyond simply crashing the system, fluctuations in the PDN can be used for surgical data corruption. When the voltage drops (a "brownout"), the signals within the FPGA travel more slowly. If the voltage drops low enough, a signal might not reach its destination before the next clock cycle begins, resulting in a timing violation.

This mechanism is the basis for "voltage hammering." In Neighbors From Hell: Voltage Attacks Against Deep Learning Accelerators on Multi-Tenant FPGAs, Andrew Boutros and Vaughn Betz show that these induced timing errors can affect physically isolated neighbors.

A particularly sophisticated application of this is the Deep-Dup framework. Developed by Adnan Siraj Rakin and Deliang Fan in Deep-Dup: An Adversarial Weight Duplication Attack Framework to Crush Deep Neural Network in Multi-Tenant FPGA, this attack uses malicious "power-plundering" circuits to cause intentional voltage drops during data transfers. These drops trigger "weight duplication" errors in a victim's Deep Neural Network (DNN). By precisely timing these faults, an attacker can destroy a model's accuracy or force it to misclassify specific inputs without ever having direct access to the model's logic.

Remote Power Side-Channels

Even if an attacker doesn't want to crash or corrupt a neighbor, they can still "listen" to them. Because the voltage on the PDN fluctuates based on the total computational load, a malicious tenant can deploy internal voltage sensors (like ring oscillators) to monitor these fluctuations.

These fluctuations act as a side-channel, leaking information about what the neighbor is calculating. This can be used to reconstruct sensitive data, such as cryptographic keys. To counter this, Mahya Morid Ahmadi and Muhammad Shafique proposed SHIELD, a defense mechanism that detects power-intensive computations in real-time and injects targeted noise. By flattening the power consumption trace, SHIELD makes it significantly harder for an attacker to distinguish the victim's signal from background noise.

Summary of the Threat Landscape

The research across these papers highlights that in the world of reconfigurable hardware, logical isolation does not equal physical security. The shared Power Distribution Network serves as a medium for:

  1. Availability Attacks: Crashing hardware via power hammering.
  2. Integrity Attacks: Corrupting computations (like DNN weights) via induced timing violations.
  3. Confidentiality Attacks: Stealing data via remote power side-channels.

As cloud providers move toward multi-tenant FPGA models, the security focus must shift from bitstream verification to active, on-chip monitoring and PDN-aware isolation.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 12 papers
  2. Version 1 · 2026-10-04 · 15 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_multi_tenant_hardware_isolation_bf148234,
  title = {multi-tenant hardware isolation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multi-tenant-hardware-isolation/rdr_bf148234acda4ceb80b8ea7c93536010},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_exploiting_shared_power_infrastructure_i_bf148234_v1,
  title = {Exploiting Shared Power Infrastructure in Multi-Tenant FPGAs},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/exploiting-shared-power-infrastructure-in-multi-tenant-fpgas/rdr_bf148234acda4ceb80b8ea7c93536010/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}