Autonomous LLM-Enhanced Adversarial Attack for Text-to-Motion

Authors: Honglei Miao, Fan Ma, Ruijie Quan, Kun Zhan, Yi Yang

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Evaluations across popular T2M models demonstrate ALERT-Motion s superiority over previous methods, achieving higher attack success rates with stealthier adversarial prompts. Sections like "4 Experiment", "4.1 Experimental Settings", "4.2 Evaluation Metrics", "4.3 Evaluation Results", and "Table 1: The results of the adversarial attacks against MDM and MLD on T2M evaluation model" confirm empirical studies with data analysis.
Researcher Affiliation Academia 1School of Information Science and Engineering, Lanzhou University 2College of Computer Science and Technology, Zhejiang University 3College of Computing and Data Science, Nanyang Technological University EMAIL
Pseudocode Yes Algorithm 1: ALERT-Motion
Open Source Code No The paper does not provide explicit access to the source code for the methodology described. It mentions using 'pretrained models from the official Git Hub repositories' for third-party tools (MLD and MDM), but not for their own ALERT-Motion.
Open Datasets Yes We select target prompt texts and target motion from the Human ML3D (H3D) (Guo et al. 2022a).
Dataset Splits No The paper mentions selecting examples for attack from the 'top 20 of the Dissimilar subset in the evaluation setup of (Petrovich, Black, and Varol 2023)' and using a 'batch size of 20, including 19 negative examples'. However, it does not provide specific training/test/validation dataset splits or detailed methodology for partitioning the data for reproducibility.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, processor types, or memory amounts) used for running its experiments.
Software Dependencies No The paper mentions using specific models and APIs like 'gpt-3.5-turbo-instruct API', 'Universal Sentence Encoder (Cer et al. 2018)', and 'GPT-2 (Radford et al. 2019)' but does not provide specific version numbers for software libraries or dependencies (e.g., PyTorch version, Python version) needed for replication.
Experiment Setup Yes We set the number of iterations as 50, the size of the prompt set as 20. We set the similarity threshold η as 0.4. Our approach involves a batch size of 20, including 19 negative examples