CS Paper Radar

Rented Oscillations: Multi-Tenant Data Center Workload Synchronization Attacks and Mitigation

Zaint A. Alexakis, Li Zhou, S. Gokul Krishnan, Charalambos Konstantinou

eess.SY cs.CR ·submitted 2026-11-18 00:49 UTC · v1

Abstract

Multi-tenant data center (DC) facilities rent computing capacity on demand, placing controllable load within reach of anyone willing to pay for it. This paper considers a stealthy adversary who rents workstations and runs deliberately unoptimized large language model (LLM) training, synchronizing their consumption into large periodic oscillations while each workload remains individually indistinguishable from benign training. The periodicity of such workloads is characterized experimentally and used to model facility consumption, showing that oscillations emerge from the interaction of concurrent workloads, whether aligned by an adversary or arising naturally. A combined small- and large-signal assessment, under a hardware-grounded threat model whose feasibility is supported by experiments on a supercomputer, locates the oscillatory weak points of the network. MW-scale oscillations are shown to saturate the generators' control limiters, amplifying their effect, while Floquet analysis reveals a lightly damped mode emerging near the forcing frequency. To counter both origins, signal-energy-based detection and mitigation are proposed, the latter re-timing the workloads to minimize the facility's oscillatory energy. The approach is validated through power hardware-in-the-loop (PHIL) experiments and real-time simulations on a surrogate model of the Saudi Arabian transmission network.

Closest papers

Nearest neighbours by meaning across the whole index.

Data Center Load Accommodation under QoS and Checkpoint-Induced Frequency Security Constraints

eess.SY ·2026-10-16 ·Ruolan Cui, Ziang Liu, Chenxi Hu, Francis Yunhe Hou
benchmark similarity 0.60
Abstract · 10 pages

The rapid growth of artificial intelligence workloads is accelerating the expansion and grid interconnection of data centers (DCs). Existing data center load accommodation (DCLA) assessments primarily focus on grid and resource adequacy and often imply that greater workload flexibility always increases accommodable capacity. However, workload flexibility has a dual effect. Workload deferral and temporal shifting can alleviate network congestion, whereas synchronized checkpointing of training workloads can cause large and rapid load reductions, threatening frequency stability. Therefore, DCLA cannot be determined solely by grid adequacy. This paper develops an optimal DCLA method that jointly accounts for workload quality of service (QoS), checkpoint-induced frequency security (FS), and grid adequacy. First, a QoS-aware model characterizes workload-dependent flexibility while ensuring service feasibility. Second, checkpoint-induced FS constraints are developed for periodic and event-triggered power variations, with key frequency-security limits reformulated in algebraic form. Third, a two-stage lexicographic model is developed to maximize DCLA and minimizes the required energy storage (ES) deployment, and then optimizes system operation under the given DC and ES siting and sizing decisions. Case studies show that the proposed method avoids a 40.3% overestimation of DCLA compared with the grid-adequacy benchmark. QoS is the primary limiting factor, while checkpoint-induced frequency-security constraints become more restrictive as training workloads increase. These findings highlight that reliable DCLA should account for both its grid-support value and the associated QoS and FS risks.

0.16

Grid-Aware Assessment of Forced Oscillations from Spatially Distributed Large Loads

eess.SY ·2026-11-17 ·Min-Seung Ko, Hao Zhu
similarity 0.59
Abstract

This paper develops a unified analytical framework for assessing forced oscillations induced by spatially distributed large-load fluctuations to determine their grid-aware fluctuation magnitudes. Emerging computational loads, such as hyperscale datacenters, present cyclic power variations with dynamic impacts depending on the waveform characteristics, forcing frequency, and source location. Our grid-aware framework represents these fluctuations using deterministic harmonic and probabilistic spectral models, and analyzes the resulting bus-wise frequency deviations based on a generic grid frequency-response model. Deterministic, risk-aware, and band-based probabilistic severity measures are further developed to obtain admissible fluctuation limits under different assumptions on frequency and phase uncertainty. We also formulate a security-constrained (SC) optimization of locational fluctuation magnitudes among existing large loads, accounting for their aggregated frequency deviation level. Case studies on the IEEE 39-bus and 14-generator Australian systems validate the proposed severity measures and admissible limit analysis. The SC optimization results further show that distributing a given amount of aggregate fluctuation toward dynamically less sensitive locations can substantially reduce system-wide oscillation severity.

0.34

Designing Grid-Aware Dynamic Specifications for Large Data Center Loads

eess.SY ·2026-09-16 ·Ashutossh Gupta, Vassilis Kekatos
similarity 0.57
Abstract · 10 pages, 8 figures, IEEE TPWRS

As data center (DC) loads increasingly penetrate the power grid, there is an urgent need for grid operators to provide clear dynamic specifications to DC owners to ensure safe grid operation. To this end, we study two salient behaviors of large language model (LLM) training loads: abrupt ramps at job initiation and termination, which induce transient frequency excursions, and sustained periodic oscillations during training, which result in oscillatory steady-state behavior. For ramping loads, we show that nodal rotor frequencies can be accurately approximated by the center-of-inertia (COI) frequency and derive analytical expressions for its nadir and rate of change of frequency (RoCoF). These expressions determine allowable combinations of ramp times and steady-state load demands satisfying prescribed frequency limits. For oscillatory loads, we derive spectral specifications on their Fourier coefficients and show that the admissible coefficient set can be approximated by a polytope. We further obtain a compact representation via its maximum-volume inscribed ellipsoid, which we show is axis-aligned. Numerical tests on the WECC 179-bus system demonstrate that the resulting specifications remain valid for higher-order nonlinear dynamics. The proposed framework provides actionable specifications for regulating the dynamic behavior of large DC loads and informing load-shaping mechanisms within the data center ecosystem.

0.50

Forced Oscillations in Power Systems Induced by Data Centers Hosting AI Workloads

eess.SY ·2026-09-23 ·Jovan Krajacic, Georgia Pierrou, Maitraya Desai, Gabriela Hug, Gustavo Valverde
similarity 0.56
Abstract · Revised and extended version of the manuscript previously posted on TechRxiv: https://www.techrxiv.org/doi/full/10.36227/techrxiv.176127304.…

Power swings in large Data Centers (DTCs) running Artificial Intelligence (AI) workloads can excite poorly damped modes in power systems. The resulting forced oscillations can lead to flicker, equipment disconnection, or blackouts. This paper investigates the risks of such load fluctuations for different system strengths and damping conditions. Using an analytical approach based on transfer functions, we identify critical DTC locations in the power grid at which load fluctuations could induce the largest forced oscillations and further characterize the harmonic spectrum of the resulting system response. Depending on the frequency and magnitude of the DTC load fluctuations, forced oscillations can become unbounded. The underlying instabilities are classified into saddle-node bifurcations of the forced periodic response and impasse-surface encounters, using Floquet multipliers and the minimum singular value of the algebraic Jacobian. Furthermore, the impact of different duty cycles and harmonic components beyond the fundamental oscillation frequency in square-wave load profiles is analyzed. Finally, the interaction of two oscillating DTCs is investigated for different locations and forcing frequencies, considering both synchronized and unsynchronized operation. The findings can help system operators to define new regulations on the maximum load fluctuations permitted for DTC facilities at specific grid locations, without negatively affecting the stability and operation of the system.

0.56

FedGuard-DC: Privacy-Preserving Federated Load Forecasting and Cyber-Attack Detection for Data-Center Loads in Transmission Systems

cs.CR ·2026-08-19 ·Md Kibria Saroare, Md Rubel Ahmed
published similarity 0.56
Abstract · Accepted on North American Power Symposium (NAPS) 2026

The rapid growth of large data-center (DC) loads is creating new challenges for power-system visibility, privacy, and cyber-physical security. System operators need accurate short-term information about these fast-varying loads, while DC operators may avoid sharing raw megawatt measurements because they can reveal sensitive workload and utilization patterns. This paper presents FedGuard-DC, a federated learning (FL) framework for privacy-preserving DC load forecasting and local false-data-injection attack (FDIA) detection. Each DC trains a dual-head model on its own measurements, where a shared encoder supports both a forecasting head and a reconstruction head. A calibrated anomaly score combines forecast residual and reconstruction error to detect corrupted measurements locally. Raw measurements and absolute MW demand remain at each DC, while only model updates, normalized forecast outputs, and scalar attack flags are shared with the global controller. Optional differential privacy and robust trimmed-mean aggregation are included to evaluate privacy-utility behavior and poisoned-client resilience. The framework is validated using phasor-domain simulation data from four large DC loads rated between 150 and 350 MW integrated into the IEEE 39-bus New England system. Results show a 0.5 s-ahead normalized forecast RMSE of 0.023-0.038 pu, compared with 0.32-0.34 pu for persistence. FedGuard-DC detects FDIA with ROC-AUC of 0.979, F1 = 0.930, and precision of 0.988, while robust aggregation reduces the poisoned-client RMSE impact from 0.042 to 0.035 pu.

0.25

Checking DER Commands Against Grid Context Catches Credentialed Attacks Static Checks Miss

cs.CR ·2026-11-09 ·Benjamin Blakely
similarity 0.56
Abstract

Authorization of distributed energy resource (DER) commands judges a command by who sent it and whether its value has been seen before. An attacker holding a legitimate credential who writes ordinary values passes both tests. We add a check that judges each DER write against the grid's operating context: the bus voltages around the target, the hour, the day- ahead price, system load, and irradiance, compared with what operators wrote under similar conditions in the feeder's own benign history. We evaluate it live on the IEEE 123-bus and 8500-node feeders, driven by a measured 2024 year of price, load, and irradiance, against nine attack scenarios. A lookup trained on the same history catches none of the writes of the four scenarios that send values already familiar on the targeted device; the context check refuses 91.7% of insider-abuse writes and 64.4% of DER-aggregation writes while denying 1.28% of benign writes. In an offline test that replays recorded moments with curtailment setpoints drawn from the benign range, issued at a positive price, it refuses 47.3% and 50.5% on the two feeders, against 18.5% and 16.0% for the same check without context; live, the few setpoints tested were mostly shallow and were mostly allowed. The context check complements the lookup.

0.40

Make the Adversary Spend More Energy: Generative Deception against Language-Model Attackers in Critical Infrastructure

cs.CR eess.SY ·2026-10-20 ·Vlad Mihal, Hosam Alamleh, Alessandro Cantelli-Forti
similarity 0.56
Abstract

Language-model attackers consume computational resources as they explore and interact with a target, creating an opportunity for cyber deception to impose costs. Whether such engagements favor the defender requires measuring both sides. We evaluate an autonomous language-model attacker against a decoy that uses a defending language model to analyze attacker commands and select defensive responses. Across 330 sessions, eight model pairs, two attacker profiles and three counter-injection levels, both sides run on identical GPUs and their energy is measured above idle. With runtime thinking disabled, the attacker-to-defender energy ratio has a 95% interval above one in 41 of 48 configurations under an allocation excluding attacker diagnostic calls, and in all 48 when those calls are included. Token ratios remain within approximately a factor of two in configurations without generation-cap truncation, showing that token counts can obscure substantial energy asymmetry. Counter-injection increases attacker tokens by approximately 9% per level without a detectable pooled change in energy consumption. Enabling thinking on both sides increases defender energy by 8--17 times and reverses the exchange for one model pair in the original campaign. We use these findings to motivate a proposed SCADA deception architecture with private physical state per identified adversary and deterministic protocol and process emulation.

0.22

Attack-Resiliency Analytics for Wide-Area Control Systems in Smart Grids

eess.SY ·2026-09-29 ·Mohammad Zakaria Haider, Prabin Mali, Nur Imtiazul Haque, Muhammad Nadeem, Sumit Paudyal +1 more
similarity 0.55
Abstract

Wide-area monitoring, protection, and control (WAMPAC) systems damp inter-area oscillations in large interconnected grids, but their reliance on synchronized PMU measurements carried over wide-area networks exposes them to false data injection (FDI) attacks. This paper presents an attack-resiliency analytics framework that formally models the coupled dynamics of the wide-area damping loop, automatic generation control, and the governor and excitation systems, and formulates the optimal stealthy FDI attack as a mixed-integer linear program (MILP). The anomaly detection model (ADM) enters as a replaceable constraint set: either a static bad-data detection (BDD) rule or a boundary learned from benign operation calibrated to a common false-positive rate. On the IEEE 39 and 118 bus systems, with reference dynamics validated on an OPAL-RT hardware-in-the-loop testbed, the optimal attack with wide-area access attains 3.3 and 8.1 times the benign objective and reaches a 0.5Hz frequency excursion more than twice as fast as an attack confined to automatic generation control. Learned boundaries reduce the attack objective by 15.5 55.0% and prevent over-frequency relay trips in all configurations, with feasible stealthy attacks remaining in every case, and residual risk tracks the width of the learned boundary rather than the detector family.

0.31