ProFlip: Targeted Trojan Attack with Progressive Bit Flips

Huili Chen, Cheng Fu, Jishen Zhao, Farinaz Koushanfar

Abstract

ProFlip is a highly efficient Trojan attack that compromises Deep Neural Networks (DNNs) by flipping a minimal number of bits in a model's memory post-deployment. By employing a progressive search strategy to identify sensitive parameters, ProFlip achieves over 94% attack success rate with significantly fewer bit flips than prior methods, highlighting a practical hardware-based threat to AI security.

The Threat of Post-Deployment Bit Flip Attacks

Traditionally, "Trojan" or "backdoor" attacks on AI models occur during the training phase, where an adversary poisons training data to embed a malicious trigger. However, this requires access to the training pipeline. Recent research has shifted focus to Bit Flip Attacks (BFA), which occur after a model is trained and deployed. By exploiting physical vulnerabilities like "Row Hammer" or using laser beams, an attacker can flip specific bits in a computer's DRAM where model weights are stored.

Overview of the Bit Flip Attack (BFA) scenario where memory tampering changes model output for triggered inputs.

The primary challenge for BFAs is the sheer scale of modern DNNs. Finding the exact bits that create a functional backdoor without degrading overall model accuracy is a "needle in a haystack" problem. Previous methods were often inefficient, requiring hundreds of bit flips and frequently targeting suboptimal model components.

The ProFlip Framework

ProFlip introduces a three-stage progressive framework designed to identify the most vulnerable bits with surgical precision. Instead of a brute-force search, it narrows the search space from the layer level down to individual bits.

  1. Salient Neurons Identification (SNI): ProFlip uses gradient-based saliency maps to identify specific neurons in the final layer that most strongly influence the attacker's target class.
  2. Trojan Trigger Generation: A specific trigger pattern (e.g., a small image patch) is crafted to stimulate these salient neurons to high activation values.
  3. Critical Bits Search (CBS): This core innovation progressively shrinks the search space: identifying the most sensitive layer, then the specific weight or bias (element level), and finally the exact bits within that element's binary representation.
The ProFlip framework flowchart, detailing the progressive search from salient neuron identification to critical bit flipping.

This search utilizes a "fitness score" that balances gradient sensitivity with the maximum possible change allowed by the model's quantization scheme.

Performance and Efficiency

The authors evaluated ProFlip using standard benchmarks including CIFAR-10, SVHN, and ImageNet across ResNet-18 and VGG-16 architectures. The results demonstrate a massive leap in efficiency:

  • Extreme Precision: To achieve a 94% attack success rate (ASR) on ResNet-18 (CIFAR-10), ProFlip required only 12 bit flips out of 88 million. For ImageNet, only 15 flips were needed.
  • 30x Improvement: Compared to the previous state-of-the-art (TBT), ProFlip reduced the required bit flips by 28× to 34×.
  • Stealthiness: The Trojaned models maintained high performance on clean data, typically seeing only a ~3% drop in accuracy, making the attack difficult to detect through standard validation.

Key Insights and Future Directions

ProFlip's sensitivity analysis revealed that internal layers (such as the penultimate convolutional layer in VGG-16) are often more susceptible to Trojan insertion than the final layer, contrary to previous assumptions. This progressive refinement avoids the computational explosion of checking every bit in massive models.

While ProFlip demonstrates a significant threat, it currently requires "white-box" knowledge of the model's architecture and weights. Furthermore, executing these flips physically via Row Hammer remains a complex engineering challenge. To counter such threats, the authors propose Weight Decomposition as a defense strategy. By mathematically splitting sensitive weight matrices into components stored in separate memory locations, the difficulty of a successful bit-flip attack can be increased by several orders of magnitude.

Summary updated

Abstract

The security of Deep Neural Networks (DNNs) is of great importance due to their employment in various safety-critical applications. DNNs are shown to be vulnerable against the Trojan attack that manipulates model parameters via poisoned training and gets activated by the pre-defined trigger during inference. In this work, we present ProFlip, the first targeted Trojan attack framework that can divert the prediction of the DNN to the target class by progressively identifying and flipping a small set of bits in model parameters. At its core, ProFlip consists of three key phases: (i) Determining significant neurons in the last layer; (ii) Generating an effective trigger pattern for the tar-get class; (iii) Identifying a sequence of susceptible bits of DNN parameters stored in the main memory (e.g., DRAM). After model deployment, the adversary can insert the Trojan by flipping the critical bits found by ProFlip using bit flip techniques such as Row Hammer or laser beams. As the result, the altered DNN predicts the target class when the trigger pattern is present in any inputs. We perform extensive evaluations of ProFlip on CIFAR10, SVHN, and ImageNet datasets with ResNet-18 and VGG-16 architectures. Empirical results show that, to reach an attack success rate (ASR) of over 94%, ProFlip requires only 12 bit flips out of 88 million parameter bits for ResNet-18 with CIFAR-10, and 15 bit flips for ResNet-18 with ImageNet. Compared to the SOTA, ProFlip reduces the number of required bits flips by 28× ∼ 34× while reaching the same or higher ASR.
BibTeX
@inproceedings{chen2021proflip,
  author = {Huili Chen and Cheng Fu and Jishen Zhao and Farinaz Koushanfar},
  title = {{ProFlip: Targeted Trojan Attack with Progressive Bit Flips}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2021},
  doi = {10.1109/iccv48922.2021.00762},
}
Figure