Bit2Watt: A Cyber-Physical Vulnerability Exploiting GPU Workloads Across Power and Computing Infrastructures

Technical Summary: Bit2Watt – A Cyber-Physical Vulnerability Exploiting GPU Workloads Across Power and Computing Infrastructures

Problem Statement
Modern data centers are increasingly integrated with renewable energy resources (DERs), such as photovoltaics (PV), creating a tightly coupled cyber-physical system. While previous research has addressed security in isolation for either power grids or computing systems, a critical gap exists regarding adversaries who operate as legitimate tenants within cloud environments. The paper posits that purely computational actions, executed via authorized GPU workloads, can be weaponized to destabilize physical power infrastructure. This vulnerability, termed Bit2Watt, exploits the intrinsic coupling between high-frequency GPU power modulations and the control dynamics of inverter-based DERs. Unlike traditional attacks that compromise sensors or communication channels, Bit2Watt operates entirely within the cyber layer, potentially bypassing routine monitoring by exploiting legitimate execution paths and concentrating distinctive behaviors in high-frequency components often missed by standard telemetry.

Methodology
The authors employ a multi-faceted approach combining theoretical modeling, simulation, and real-world experimentation to validate the Bit2Watt risk and its feedback loop, Watt2Bit.

1. Attack Vectors (Cyber Layer): The paper proposes two methods for inducing power fluctuations without elevated privileges:

   • Synthetic Workload Modulation Attack (SWMA): A custom CUDA kernel that toggles between high-intensity compute modes (saturating SMs) and idle states via a user-space control flag, achieving modulation frequencies up to 6,000 Hz.
   • LLM Training Modulation Attack (LTMA): Embedding modulation logic into standard Large Language Model training pipelines (e.g., PyTorch, DeepSpeed) by manipulating hyperparameters or batch schedules. This camouflages malicious signatures as legitimate computational noise.

2. Physical Modeling and Analysis:

   • Impedance-Based Modeling: The system is modeled using Thévenin and Norton equivalents, treating the GPU cluster as a Constant Power Load (CPL) with negative incremental resistance (NIR) and DERs as controlled current sources with frequency-shaped output impedances.
   • Stability Analysis: The authors analyze how GPU-induced current fluctuations interact with DER control loops (PLL, voltage, and current controllers). They demonstrate that high-frequency modulations can exceed the response bandwidth of DER controls, leading to harmonic amplification and negative damping ratios.

3. Validation:

   • Hardware Experiments: Real-world tests were conducted on modified workstations with NVIDIA GPUs (RTX 3090, V100, A100) connected to a local grid testbed featuring PV inverters and UPS systems. Measurements captured current/voltage waveforms and harmonic spectra.
   • Simulations: Extensive simulations were performed using Simulink on a 1-MW local system and the IEEE 118-bus and 9241-bus transmission networks to assess cascading failure risks under varying DER penetration levels (up to 90%).

Key Contributions

1. Formalization of Bit2Watt: The paper defines a new class of cyber-physical threats where computational processes act as controllable excitation sources interacting with power system dynamics, establishing a link between workload orchestration and grid-level instability.
2. Novel Attack Techniques: The development of SWMA and LTMA, which achieve high-frequency power modulations (up to 6 kHz) using only user-level access, effectively bypassing cloud-side monitoring frameworks.
3. Impedance-Based Vulnerability Analysis: A rigorous analytical framework quantifying how GPU power modulations interact with inverter-dominated grids, revealing that high-frequency fluctuations can induce negative damping and system instability.
4. Watt2Bit Feedback Loop: The identification of a self-reinforcing cycle where power quality degradation triggers protection mechanisms (over-temperature, over-current), leading to Denial of Service (DoS) and potential covert information exfiltration via EMI side channels.
5. Empirical Validation: Quantitative evidence from both simulations and real-world experiments demonstrating that manipulating a small fraction of GPUs (e.g., 1,000 units in a 1-MW system) can drastically degrade power quality (THD up to 46.8%) and trigger cascading failures in large-scale networks.

Results

• Power Quality Degradation: In a simulated 1-MW local grid with 90% DER penetration, manipulating 1,000 GPUs raised the Total Harmonic Distortion (THD) of current to 46.8% and reduced the system damping ratio to -0.27, indicating an unstable regime.
• System Instability: The negative incremental resistance of GPU clusters, combined with high DER penetration, reduces stability margins. Simulations on the IEEE 118-bus system showed that attacks can cause frequency bifurcation, splitting the network into coherent oscillatory clusters.
• Cascading Failures: In a modified 9241-bus European transmission network simulation, a localized disturbance (2% of total load) triggered a 13-stage cascade, resulting in the disconnection of 1,238 lines and an 81% load loss.
• Watt2Bit Impacts:
  • DoS: Increased RMS current and thermal stress from harmonics can trip protection relays (e.g., IEC 60255 IDMT), causing abrupt shutdowns of GPU servers.
  • Exfiltration: Experiments demonstrated that modulating GPU loads could encode binary data into EMI side channels (using FSK at 2 kHz and 200 Hz), achieving >99% decoding accuracy.
• Stealth: Standard cloud monitoring tools (NVML, RAPL, BMC) with low sampling rates (1 Hz to 1 kHz) fail to detect the high-frequency signatures of Bit2Watt, rendering the attack highly covert.

Significance
The paper argues that the convergence of computing and power infrastructures necessitates a paradigm shift in security. The Bit2Watt vulnerability highlights that legitimate computational workloads can be exploited to destabilize physical grids, particularly in high-DER scenarios where system inertia is low and control loops are complex. The findings underscore the urgent need for cross-layer defenses that jointly consider workload scheduling, power electronics dynamics, and grid stability. The authors emphasize that isolated security measures in either the cyber or physical domain are insufficient to mitigate this closed-loop risk, calling for integrated monitoring and protection strategies for future renewable-integrated data centers.