- Rating
- 3(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper presents a bit-flip attack framework for targeted bit-flip attacks on LLMs with the help of a keyword-focused attack loss. The paper evaluates the framework across a myriad of tasks and models across different families. The paper demonstrates that 1-9 weight bit-flips are enough to bring about targeted output changes in LLMs with less collateral degradation.
Critical strengths
- The paper identifies targeted bit-flip attack (BFAs) as an important problem requiring investigation, as most of the recent bit-flip attack frameworks have focused on untargeted bit-flip attack. It is important to assess whether it is possible to perform targeted BFAs as it could lead to LLMs generating misleading outputs with high confidence.
- Similar to the observations for untargeted BFAs, targeted BFAs also require 1-9 bit-flips, indicating that the effort to execute a targeted BFA could be similar.
Critical weaknesses
- The keyword-focused attack loss is presented as a core contribution of the paper. However, the attack loss is described only in a subsection without much theoretical basis. There lacks a thorough analysis regarding why the particular masking strategy in the proposed loss produces a better targeted control compared to cross-entropy of the complete response? The presentation of this section seems to be empirical at best. Merely defining the different symbols used in the definition does not equate to providing a theoretical basis for the loss. Furthermore, the current form of contributions state keyword-focused attack loss and AuxUtility Score to be separate novel things. However, the general concept of using a gradient-based sensitivity for AuxUtilityScore, in one form of another, has been used in past work such as the 1P-DNL scoring for layer selection in Galil et. al. (https://arxiv.org/pdf/2502.07408). Thus, I would interpret the novelty to be incremental.
- The reproducibility of the paper is incomplete. The authors have not discussed how the baselines have been set up so that the comparison is apples-to-apples. For example, the GenBFA algorithm is mainly for an untargeted attack, thus a targeted attack will always fail if the framework is set up as-is. Thus, it is not known how faithful is the comparison with the baselines presented. This setup is also absent from the appendix.
Questions for authors
- Can the authors provide a mathematical explanation and sensitivity analysis for the use of keyword-focused loss with the AuxUtilityScore, in terms of why the sensitivity proxies in works such as AttentionBreaker (TMLR 2025) or 1P-DNL cannot be used?
- How many target and auxiliary samples are used in the gradient computation and candidate evaluation? Is there any kind of sensitivity of the number of attack samples to the attack success rate, bit-flip budget, or the computational overhead?
- What proportions of successful bit flips affect the sign, exponent, and mantissa fields for BF16, and how does this distribution compare with the selected bit positions under INT8 quantization?
- Can the authors elaborate on the experimental setup for the baselines that are evaluated to show that the presented TFL framework is better?
- Can the authors provide a reasoning why the use of keyword-focused loss is not affecting the model perplexity, thereby retaining the model’s language modeling capabilities? The targeted changes explained in the Appendix F.1 seem to indicate that the responses remain comprehensible, thus indicating the model perplexity could be less.
5 comments
Rebuttal to reviewer zppq Cont.
Additional comparison using the same attack objective (loss)
To test the alternative candidate ranking and search procedures directly, we evaluated AttentionBreaker (GenBFA) and 1P-DNL on the two main QA targets using our Keyword-Focused Attack Loss. Neither method achieves either target: both obtain 0/2 targeted success, despite reducing the attack loss.
| Method | Bit flips | Target success | Initial loss | Final loss | DROP F1 | GSM8K accuracy | TriviaQA accuracy |
|---|---|---|---|---|---|---|---|
| AttentionBreaker | 209,715 | 0/2 | 1.9474 | 1.6000 | 0.6781 | 0.8636 | 0.4524 |
| 1P-DNL | 30 | 0/2 | 1.9474 | 1.8845 | 0.7759 | 0.9545 | 0.5238 |
AttentionBreaker fails to produce the specified answers even with 209,715 flips, while 1P-DNL fails with 30 flips. Neither method achieves our attack objective in this evaluation.
In conclusion, our keyword-focused loss and AuxUtilityScore are important components of TFL and differ from the sensitivity proxies used in AttentionBreaker and 1P-DNL. Even when evaluated under the same attack objective, both methods fail to achieve the targeted manipulation. This further supports the contribution and effectiveness of TFL's design choices.
W2. Target and auxiliary sample counts and sensitivity
The main QA setting uses two target questions and four related benign questions, with two benign questions per target. Section 6.2 examines the effect of targeting five attack samples simultaneously: the five-target joint attack requires 26 flips compared to the original two-target attack which requires 4 flips, and it incurs greater collateral degradation than the original two-target attack. As runtime increase linearly with number of flips, there is more computation overhead for the five-target joint attack.
W3. BF16 and INT8 bit-position distributions
Across the main experiments, all BF16 flips affect exponent bits: 33.3% occur at b9 and 66.7% at b10, with no sign- or mantissa-bit flips. All are 0→1 flips, increasing the magnitude of normal finite weights by factors of 16 and 256, respectively. Under signed two’s-complement INT8 quantization, all flips affect b7 (MSB): 85.7% are 0→1, decreasing the stored integer value by 128, and 14.3% are 1→0, increasing it by 128. The corresponding dequantized weight changes depend on the quantization scale. So BF16 flips change the weight magnitude more than INT8 flips in most cases.
W4. Baseline experimental setup
We compare existing BFA methods under (1) their original attack objectives and (2) our targeted attack objective through adapted baselines and search ablations.
First, to our knowledge, TFL is the first bit-flip attack framework for prompt-specific targeted response manipulation in generative LLMs while limiting collateral degradation. Section 5.1 tests whether existing methods provide this capability under their original objectives. They fail to produce the attacker-specified responses, whereas TFL succeeds, supporting this contribution.
Second, with our Keyword-Focused Attack Loss, AttentionBreaker (GenBFA) and 1P-DNL both achieve 0/2 targeted success (W1). Appendix B compares gradient-based search with ImpactScore under the same targeted objective; the gradient-based (used by PrisonBreak and SilentStriker) alternatives fail within the allowed budget. These comparisons support TFL's framework is better.
W5. Perplexity and language-modeling performance
As discussed in W1, AuxUtilityScore evaluates the relative changes in target loss and auxiliary utility after temporarily applying each candidate flip. When WikiText is used, auxiliary utility is defined as negative language-modeling loss, so candidate selection explicitly accounts for language-modeling performance. Thus, limiting increases in model perplexity is primarily the role of AuxUtilityScore, while the Keyword-Focused Attack Loss promotes attacker-specified target tokens.
For example, the attacked model generates the following summary, replacing the original detail “four days later” with the attacker-selected “two days later”:
A stray dog in Washington State, named Theia, miraculously survived after being struck by a car, attacked with a hammer, and buried in a field, only to be discovered two days later. Supporters and a local fund have rallied around Theia, raising over $10,000 for her surgeries, as she continues to recover from severe injuries including a dislocated jaw and damaged sinuses. Theia’s story highlights both the resilience of animals and the power of community compassion in aiding those facing extreme adversity.
This example demonstrates targeted factual manipulation within a coherent response.
Rebuttal to reviewer zppq Cont.
Concluding remarks
We clarify how our Keyword-Focused Attack Loss and AuxUtilityScore differ from the sensitivity proxies used in other methods and reaffirm their importance in achieving targeted manipulation. We also clarify the sample settings, baseline setup, bit-position distributions, and how AuxUtilityScore accounts for language-modeling performance. Together, this evidence reinforces TFL's contribution to prompt-specific targeted response manipulation in generative LLMs through the Keyword-Focused Attack Loss and AuxUtilityScore, while limiting collateral degradation. We respectfully ask the reviewer to reconsider the assessment and raise the score based on these clarifications and additional results.
Rebuttal to reviewer zppq
Thank you for recognizing the importance of targeted bit-flip attacks and for your detailed questions about the method, baseline setup, and preservation of model utility.
W1. Keyword-focused loss, Aux Utility Score, and sensitivity proxies
We thank the reviewer for the question. We would like to clarify a distinction between the attack objective, sensitivity-based candidate ranking, and final bit-flip selection. The sensitivity proxies in AttentionBreaker and 1P-DNL are not direct equivalents of AuxUtilityScore. They belong to different search procedures and do not explicitly provide the auxiliary-utility trade-off addressed by AuxUtilityScore.
Role of the Keyword-Focused Attack Loss and AuxUtilityScore
The Keyword-Focused Attack Loss is defined in Section 4.2, Eq. (1). It promotes attacker-specified target tokens and supplies the gradients used for candidate search. TFL then uses a two-stage selection procedure:
Stage 1. Candidate search: ImpactScore, defined in Appendix B, Eq. (5), combines the attack-loss gradient magnitude with the maximum permitted bit-induced weight perturbation under the range constraint in Eq. (4). It populates the Top-k candidate queue. Stage 2. Final flip selection: AuxUtilityScore, defined in Section 4.3, Eq. (2), evaluates each candidate through a temporary flip and re-ranks the queue using the relative change in target loss and auxiliary utility. The candidate with the minimum score is selected.
These stages serve different purposes. ImpactScore estimates the magnitude of a candidate's local influence. AuxUtilityScore evaluates the relative changes in target loss and auxiliary utility after temporarily applying each candidate flip. A large sensitivity score alone does not establish that a discrete bit flip reduces the target loss while preserving benign performance.
Differences from AttentionBreaker and 1P-DNL
AttentionBreaker’s search procedure can be decomposed into four stages: (1) ranking weights within each candidate matrix, (2) assessing matrix sensitivity through temporary bit perturbations and selecting a sensitive matrix, (3) constructing a candidate weight pool within that matrix, and (4) using GenBFA to optimize the subset of weights whose bits are flipped. Its final selection considers attack loss and flip count, rather than the auxiliary-utility change used by AuxUtilityScore. In contrast, 1P-DNL computes a saliency ranking from one forward/backward pass on random inputs and directly selects the top-k weights for bit-flips (one stage). It has no separate auxiliary-aware re-ranking stage. Thus, neither method contains a component directly equivalent to Eq. (2), although both can be adapted as alternative search & selection methods.
Keyword-Focused Attack Loss vs. standard full-sequence loss
We evaluated standard full-sequence loss on LLaMA-3.1-8B-INSTRUCT in BF16 using the main experiment's president and mountain targets. It fails to generate the specified target responses after 50 bit flips, with post-attack benchmark scores falling to near zero. In comparison, TFL with the Keyword-Focused Attack Loss succeeds with only two flips in Table 1 and maintains substantially higher benchmark performance.
Full-sequence loss distributes optimization across the entire response, whereas the Keyword-Focused Attack Loss concentrates it on attacker-specified target tokens. Bit flips differ from general training updates in that each update is limited to a narrow scope of "knowledge". A keyword-focused loss is more effective for targeted manipulation where standard loss may be dominated by other tokens in the sequence.
AuxUtilityScore vs without AuxUtilityScore
Section 6.1 and Table 4 examine the effectiveness of AuxUtilityScore, without it TFL can not achieve the target manipulation or required more flips and more collateral degradation.