Sponge Examples: Energy-Latency Attacks on Neural Networks

2020/06/05 by Ilia Shumailov, Shumailov, Ilia, Yiren Zhao +9 · 1 voice · 27 citations
Computer Science · Mathematics · #Computation and Language (cs.CL) #Cryptography and Security (cs.CR) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning (stat.ML) #cs.CL #cs.CR #cs.LG #stat.ML

paper · pdf · doi:10.48550/arxiv.2006.03463

Accepted at 6th IEEE European Symposium on Security and Privacy (EuroS&P)

arxiv published 2020/06/05 · arxiv created 2021/05/12 · arxiv updated 2021/05/13

Abstract

The high energy costs of neural network training and inference led to the use of acceleration hardware such as GPUs and TPUs. While this enabled us to train large-scale neural networks in datacenters and deploy them on edge devices, the focus so far is on average-case performance. In this work, we introduce a novel threat vector against neural networks whose energy consumption or decision latency are critical. We show how adversaries can exploit carefully crafted \boldsymbolsponge~\boldsymbolexamples, which are inputs designed to maximise energy consumption and latency. We mount two variants of this attack on established vision and language models, increasing energy consumption by a factor of 10 to 200. Our attacks can also be used to delay decisions where a network has critical real-time performance, such as in perception for autonomous vehicles. We demonstrate the portability of our malicious inputs across CPUs and a variety of hardware accelerator chips including GPUs, and an ASIC simulator. We conclude by proposing a defense strategy which mitigates our attack by shifting the analysis of energy consumption in hardware from an average-case to a worst-case perspective.

Cited by

• ThinkTrap: Denial-of-Service Attacks against Black-box LLM Services via Infinite Thinking
• Toward a Safe Internet of Agents
• An Image Is Worth Ten Thousand Words: Verbose-Text Induction Attacks on VLMs
• LoopLLM: Transferable Energy-Latency Attacks in LLMs via Repetitive Generation
• StealthAttack: Robust 3D Gaussian Splatting Poisoning via Density-Guided Illusions
• Known By Their Actions: Fingerprinting LLM Browser Agents via UI Traces
• Adversarial News and Lost Profits: Manipulating Headlines in LLM-Driven Algorithmic Trading
• Driving up Inference Energy on SNNs: Per-Sample and Universal Sponge Attacks
• Sy-FAR: Symmetry-based Fair Adversarial Robustness
• LLM in the Middle: A Systematic Review of Threats and Mitigations to Real-World LLM-based Systems
• On Surjectivity of Neural Networks: Can you elicit any behavior from your model?
• Hidden Tail: Adversarial Image Causing Stealthy Resource Consumption in Vision-Language Models
• Security Analysis of ChatGPT: Threats and Privacy Risks
• CP-FREEZER: Latency Attacks against Vehicular Cooperative Perception
• BadReasoner: Planting Tunable Overthinking Backdoors into Large Reasoning Models for Fun or Profit
• Resource Consumption Red-Teaming for Large Vision-Language Models
• Securing AI Systems: A Guide to Known Attacks and Impacts
• LingoLoop Attack: Trapping MLLMs via Linguistic Context and State Entrapment into Endless Loops
• Excessive Reasoning Attack on Reasoning LLMs
• Efficiency Robustness of Dynamic Deep Learning Systems
• Reality Check: A New Evaluation Ecosystem Is Necessary to Understand AI's Real World Effects
• PD³F: A Pluggable and Dynamic DoS-Defense Framework Against Resource Consumption Attacks Targeting Large Language Models
• BitHydra: Towards Bit-flip Inference Cost Attack against Large Language Models
• The Boy Who Cried Wolf: Adversarial Misclassification of Safe Inputs as Unsafe in Multimodal Guardrails
• Sponge Attacks on Sensing AI: Energy-Latency Vulnerabilities and Defense via Model Pruning
• ThreMoLIA: Threat Modeling of Large Language Model-Integrated Applications
• Safety in Large Reasoning Models: A Survey
• EO-VLM: VLM-Guided Energy Overload Attacks on Vision Models

Discussions

• Sponge Examples: Energy-Latency Attacks on Neural Networks [hn, 2 points, 0 comments]

Related

• Energy-Latency Attacks: A New Adversarial Threat to Deep Learning
• Energy-Latency Attacks to On-Device Neural Networks via Sponge Poisoning
• Sponge Attacks on Sensing AI: Energy-Latency Vulnerabilities and Defense via Model Pruning
• Driving up Inference Energy on SNNs: Per-Sample and Universal Sponge Attacks
• Energy Backdoor Attack to Deep Neural Networks