Abstract Recent studies have indicated that large language models (LLMs) remain susceptible to adversarial attacks, despite enhanced robustness through the chain-of-thought (CoT) capability. However, this capability also introduces the potential for more covert and effective adversarial attack methods. This paper proposes a CoT Transfer Adversarial attack framework (CTTA) for general LLMs. Initially, we utilize a pre-trained model based on the transformer architecture and fine-tune it on various tasks to serve as a surrogate model. Subsequently, different levels of adversarial attack algorithms are utilized, and the generated adversarial samples are used as transfer samples. A thought chain-based adversarial transfer attack framework is constructed using transfer samples and thought chain techniques. Finally, various indicators are utilized to assess the performance of the general LLMs in response to this attack. The results demonstrate that the attack framework surpasses current state-of-the-art research. Numerous experiments on LLMs with varying performance and parameter sizes have validated the effectiveness, stability, and generalizability of this attack. The model’s error response and the superiority of this attack are thoroughly examined using attention by gradient technology, confirming the security threats posed by LLMs when leveraging CoT capability. This has significant implications for enhancing the security and robustness of LLMs.

Thread-Level Speculation (TLS) is a thread-level automatic parallelization technique to accelerate sequential programs on multi-core. Thread partition is a core step for this technique, so how to automatically and effectively partition an unknown program is a key to improve the efficiency of this technique. In order to solve this problem, this paper proposes a Back Propagation Neural Network based threading partition approach(TPoBP). This approach is used to study the implicit knowledge of partition in the sample set to guide the partition for unknown programs. The knowledge in the sample is composed of the characteristics of the sample and the partition scheme, which are used as the input and output of the network to train the network until the specified accuracy is reached. During validation period, the trained network makes use of profiling information (obtained during pre-execution) of a validation program as input, and runs to obtain the predicted partition scheme for the validation program. Experimental results show that TPoBP can effectively predict the partition schemes of validation programs, and average prediction accuracy almost reaches 0.7. Moreover, these predicted schemes are further used to guide partition for validation programs, and Olden benchmarks reach a maximum 11.8% speedup improvement. Experiments demonstrate that the model proposed by this paper is effective to predict partition scheme for unseen programs.

Indiscriminate data poisoning attacks are highly effective against unsupervised learning. However, recent studies show that contrastive learning is also susceptible to data poisoning attacks. As a form of data poisoning attack, the attacker adds poison to the clean pre-training dataset. This article proposes IDPA, an indiscriminate data poisoning attack targeting the encoder in contrastive learning. where the attacker’s goal is to directly poison the pre-trained encoder. The feature vectors of any clean sample and the attacked sample from the attacker will exhibit high similarity, causing the downstream classifier to misclassify the clean sample as the samples designated by the attacker. Therefore, this article formulates IDPA as a dual optimization problem and defines two loss functions: the attack effectiveness loss and the model utility loss. These losses are associated with effectively poisoning the pre-trained encoder and maintaining the accuracy of the downstream classifier, respectively. During training, the attack affects the contrastive learning algorithm and predictions are made on multiple datasets. Experimental results show that the attack success rate of 92%. This article evaluates the effectiveness of IDPA on the CLIP dataset released by OpenAI, with attack success rate of 88%.

We use multiple satellite observations and numerical simulation to investigate the upper ocean response to Typhoon Kalmaegi in 2014 in the South China Sea (SCS). In this study, significant sea surface temperature (SST) decreasing is observed, which is caused by typhoon-induced vertical mixing and upwelling. The maximum SST cooling is 2 °C on the right of the typhoon track since inertial currents rotate in the same direction as the surface wind vectors. We also find an interesting phenomenon that salinity decreases ranging between 0.3 and 0.6 psu on the left side of the typhoon track. We use numerical simulations and in-situ observations to further confirm that salinity reduction is caused by heavy rainfall.