Emotion-Based Backdoor Attack Threatens LLM Security
Researchers have identified a novel backdoor vulnerability in large language model fine-tuning that exploits emotional triggers. According to arxiv.org, the attack method called “Paraesthesia” uses emotion as a backdoor trigger rather than traditional token-level manipulation, achieving “an attack success rate of around 99% across both task types and four different models, while maintaining the clean utility of the models.”
The research observes that “emotion is not directly linked to individual words, but functions as an overall stylistic factor through tone,” making it more difficult to detect than static triggers. The attack works by mixing emotionally-triggered samples into clean training data during fine-tuning, causing the model to generate predefined harmful responses when encountering emotional inputs during inference.
Carbon-Aware and Edge-Optimized LLM Inference
Separate research addresses environmental and hardware constraints in LLM deployment. According to arxiv.org, Green-Aware Routing (GAR) introduces “a constrained multi-objective optimization framework that minimizes per-request CO2 emissions subject to explicit accuracy floors and p95-latency service-level objectives.”
For memory-constrained devices, researchers developed CATS (Cascaded Adaptive Tree Speculation), which according to arxiv.org “can achieve a wall-clock speedup of up to 5.08x with no degradation in generation quality” on edge devices while keeping memory footprint equal to the target model alone.