Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

Backdoor Attacks in Token Selection of Attention Mechanism

Authors: Yunjuan Wang, Raman Arora

ICML 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically validate our theoretical findings using synthetic datasets.
Researcher Affiliation Academia 1Department of Computer Science, Johns Hopkins University, Baltimore, USA.
Pseudocode No The paper describes mathematical proofs and theoretical analysis but does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any concrete access information for source code, such as a repository link, an explicit code release statement, or mention of code in supplementary materials.
Open Datasets Yes We empirically validate our theoretical findings using synthetic datasets. We adopt the dirty-label backdoor attack setup defined in Section 3. Standard and poisoned signal vectors are constructed from orthogonal basis directions: µ1 = µ e1, µ2 = µ e2, µ1 = α µ e3, µ2 = α µ e4. We designate the first |R| tokens as relevant and the last |P| tokens as poisoned for the poisoned data. We generate n = 20 training samples, along with 1000 standard test samples and 1000 poisoned test samples. Noise vectors are drawn from a standard multivariate Gaussian with covariance Σ = Id, yielding Tr(Σ) = d.
Dataset Splits Yes We generate n = 20 training samples, along with 1000 standard test samples and 1000 poisoned test samples.
Hardware Specification No The paper does not provide specific hardware details such as GPU or CPU models, processor types, or memory amounts used for running the experiments.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers, such as programming languages, libraries, or solvers, which are needed to replicate the experiment.
Experiment Setup Yes The token length is set to T = 8, dimension d = 4000, with |R| = 1 and |P| = 1. A single-head self-attention transformer is trained using gradient descent with step size η = 0.001 for τ0 = 10K iterations. Additional results are provided in Appendix B. [...] We run τ0 = 1K iterations with step size η = 0.01.