Search for: All records

The remarkable performance of large language models (LLMs) in generation tasks has enabled practitioners to leverage publicly available models to power custom applications, such as chatbots and virtual assistants. However, the data used to train or fine-tune these LLMs is often undisclosed, allowing an attacker to compromise the data and inject backdoors into the models. In this paper, we develop a novel inference time defense, named CLEANGEN, to mitigate backdoor attacks for generation tasks in LLMs. CLEANGEN is a lightweight and effective decoding strategy that is compatible with the state-of-the-art (SOTA) LLMs. Our insight behind CLEANGEN is that compared to other LLMs, back doored LLMs assign significantly higher probabilities to tokens representing the attacker-desired contents. These discrepancies in token probabilities enable CLEANGEN to identify suspicious tokens favored by the attacker and replace them with tokens generated by another LLM that is not compromised by the same attacker, thereby avoiding generation of attacker-desired content. We evaluate CLEANGEN against five SOTA backdoor attacks. Our results show that CLEANGEN achieves lower attack success rates (ASR) compared to five SOTA baseline defenses for all five backdoor attacks. Moreover, LLMs deploying CLEANGEN maintain helpfulness in their responses when serving benign user queries with minimal added computational overhead. 

We introduce Boundless, a photo-realistic synthetic data generation system for enabling highly accurate object detection in dense urban streetscapes. Boundless can replace massive real-world data collection and manual groundtruth object annotation (labeling) with an automated and configurable process. Boundless is based on the Unreal Engine 5 (UE5) City Sample project with improvements enabling accurate collection of 3D bounding boxes across different lighting and scene variability conditions. We evaluate the performance of object detection models trained on the dataset generated by Boundless when used for inference on a real-world dataset acquired from medium-altitude cameras. We compare the performance of the Boundless-trained model against the CARLA-trained model and observe an improvement of 7.8 mAP. The results we achieved support the premise that synthetic data generation is a credible methodology for training/fine-tuning scalable object detection models for urban scenes. 

Acer(Sapindaceae) is a major genus of broadleaf trees dominating deciduous forests in the Northern Hemisphere, with Asia exhibiting the highest species diversity. Many economically importantAcerspecies are cultivated for ornamental or timber purposes.Acerpowdery mildew, caused by fungi in the tribeCystotheceae, poses significant global economic and ecological threats. The pathogenicity spectrum remains unclear due to taxonomic uncertainties in its primary causal genera,SawadaeaandTakamatsuella. This study presents a comprehensive phylogenetic-taxonomic analysis of the two genera across East Asia, Europe, and North America. Using 75 ITS and 58 28S rDNA newly obtained sequences, we resolved 12Sawadaeaspecies and oneTakamatsuellaspecies into nine monophyletic clades, revealing marked cryptic diversity (three new species:S. acerina,S. aceris-arguti,S. taii) and two paraphyletic groups (S. bifida/S. negundinis). Taxonomic revisions include:S. bicornissplit into twoformae(f. bicornisandf. polyphaga f. nov.) with distinct host preferences;S. tulasnei(sensu stricto) restricted to Europe/North America, invalidating previous Asian records;S. nankinensisandS. koelreuteriaeform two basal lineages. Phylogenetic positioning confirmedTakamatsuellaas a distinct genus sister toSawadaea, supported by an ITS1 26 bp deletion. Host specificity analysis revealed narrow host ranges (primarilyAcer) with two evolutionary host expansions toKoelreuteria,Aesculus, andLiquidambar. This study also newly describes the asexual morphs of four species (S. aesculi,S. bifida,S. bomiensisandS. kovaliana) and establishes a molecular framework for disease management through clarified phylogeny and taxonomy. Our findings provide critical insights into fungal evolution, host-pathogen interactions, and strategies for mitigating powdery mildew impacts in forest ecosystems. 

We have designed a new filter pack array to measure angular variations in x-ray spectra during a single shot. The filter pack was composed of repeating identical columns of aluminum and copper filters of varying thicknesses. These columns were located at different positions to measure the spectrum at each corresponding angle. This array was utilized in an experiment to measure the energy evolution of betatron x rays in a laser wakefield accelerator by curving the wakefield with a transverse density gradient, streaking the x rays across the array in front of an x-ray charge-coupled device (CCD) camera. After subtracting the background and “flattening” the image to remove spatial nonuniformities, a critical energy was calculated for each position that produced the best agreement with the measured signal. There was a clear change in critical energy with angle, shedding light on the dynamics of the electrons that traveled through the accelerator. These angles correspond to distinct emission times, covering a timescale of tens of picoseconds. The filter pack was capable of recovering these angular details without the impact of errors introduced by shot-to-shot variability.