As text-to-image (T2I) diffusion models become increasingly deployed in real-world applications, concerns about backdoor attacks have gained significant attention. Prior works largely focus on models with a single text encoder, leaving recent multi-encoder architectures, e.g., Stable Diffusion 3 (SD 3), underexplored. In this paper, we study SD 3, which uses three distinct text encoders, and systematically analyze its text-encoder-based backdoor vulnerabilities. To understand each text encoder’s role in backdoor attacks, we define four target categories spanning from global-level to entity-level semantics and identify the minimal encoder subset required for each attack type. Furthermore, we propose Multi-Encoder Lightweight aTtacks (MELT) and validate MELT’s effectiveness and parameter efficiency. MELT matches full encoder fine-tuning attack performance while using fewer than (0.2%) of its parameters, demonstrating that effective backdoors can be implanted under limited parameter budgets.

The spread of disinformation poses a global threat to democratic societies, necessitating robust and scalable Automated Fact-Checking (AFC) systems. The AVeriTeC Shared Task Challenge 2024 offers a realistic benchmark for text-based fact-checking methods. This paper presents Information-Retrieving Fact-Checker (InFact), an LLM-based approach that breaks down the task of claim verification into a 6-stage process, including evidence retrieval. When using GPT-4o as the backbone, InFact achieves an AVeriTeC score of 63% on the test set, outperforming all other 20 teams competing in the challenge, and establishing a new strong baseline for future text-only AFC systems. Qualitative analysis of mislabeled instances reveals that InFact often yields a more accurate conclusion than AVeriTeC’s human-annotated ground truth.