Rishi Sunak's advisers are trying to thrash out an agreement among world leaders on a statement warning about the risks of artificial intelligence as they finalise the agenda for the AI safety summit next month. Downing Street officials have been touring the world talking to their counterparts from China to the EU and the US as they work to agree on words to be used in a communique at the two-day conference. But they are unlikely to agree a new international organisation to scrutinise cutting-edge AI, despite interest from the UK in giving the government's AI taskforce a global role. Sunak's AI summit will produce a communique on the risks of AI models, provide an update on White House-brokered safety guidelines and end with "like-minded" countries debating how national security agencies can scrutinise the most dangerous versions of the technology. The possibility of some form of international cooperation on cutting-edge AI that can pose a threat to human life will also be discussed on the final day of the summit on 1 and 2 November at Bletchley Park, according to a draft agenda seen by the Guardian.

Graph neural networks (GNNs) are vulnerable to adversarial perturbations, including those that affect both node features and graph topology. This paper investigates GNNs derived from diverse neural flows, concentrating on their connection to various stability notions such as BIBO stability, Lyapunov stability, structural stability, and conservative stability. We argue that Lyapunov stability, despite its common use, does not necessarily ensure adversarial robustness. Inspired by physics principles, we advocate for the use of conservative Hamiltonian neural flows to construct GNNs that are robust to adversarial attacks. The adversarial robustness of different neural flow GNNs is empirically compared on several benchmark datasets under a variety of adversarial attacks. Extensive numerical experiments demonstrate that GNNs leveraging conservative Hamiltonian flows with Lyapunov stability substantially improve robustness against adversarial perturbations.

Cybersecurity of Industrial Cyber-Physical Systems is drawing significant concerns as data communication increasingly leverages wireless networks. A lot of data-driven methods were develope for detecting cyberattacks, but few are focused on distinguishing them from equipment faults. In this paper, we develop a data-driven framework that can be used to detect, diagnose, and localize a type of cyberattack called covert attacks on networked industrial control systems. The framework has a hybrid design that combines a variational autoencoder (VAE), a recurrent neural network (RNN), and a Deep Neural Network (DNN). This data-driven framework considers the temporal behavior of a generic physical system that extracts features from the time series of the sensor measurements that can be used for detecting covert attacks, distinguishing them from equipment faults, as well as localize the attack/fault. We evaluate the performance of the proposed method through a realistic simulation study on a networked power transmission system as a typical example of ICS. We compare the performance of the proposed method with the traditional model-based method to show its applicability and efficacy.

We introduce new jailbreak attacks on vision language models (VLMs), which use aligned LLMs and are resilient to text-only jailbreak attacks. Specifically, we develop cross-modality attacks on alignment where we pair adversarial images going through the vision encoder with textual prompts to break the alignment of the language model. Our attacks employ a novel compositional strategy that combines an image, adversarially targeted towards toxic embeddings, with generic prompts to accomplish the jailbreak. Thus, the LLM draws the context to answer the generic prompt from the adversarial image. The generation of benign-appearing adversarial images leverages a novel embedding-space-based methodology, operating with no access to the LLM model. Instead, the attacks require access only to the vision encoder and utilize one of our four embedding space targeting strategies. By not requiring access to the LLM, the attacks lower the entry barrier for attackers, particularly when vision encoders such as CLIP are embedded in closed-source LLMs. The attacks achieve a high success rate across different VLMs, highlighting the risk of cross-modality alignment vulnerabilities, and the need for new alignment approaches for multi-modal models.

Machine learning (ML) - based software systems are rapidly gaining adoption across various domains, making it increasingly essential to ensure they perform as intended. This report presents best practices for the Test and Evaluation (T&E) of ML-enabled software systems across its lifecycle. We categorize the lifecycle of ML-enabled software systems into three stages: component, integration and deployment, and post-deployment. At the component level, the primary objective is to test and evaluate the ML model as a standalone component. Next, in the integration and deployment stage, the goal is to evaluate an integrated ML-enabled system consisting of both ML and non-ML components. Finally, once the ML-enabled software system is deployed and operationalized, the T&E objective is to ensure the system performs as intended. Maintenance activities for ML-enabled software systems span the lifecycle and involve maintaining various assets of ML-enabled software systems. Given its unique characteristics, the T&E of ML-enabled software systems is challenging. While significant research has been reported on T&E at the component level, limited work is reported on T&E in the remaining two stages. Furthermore, in many cases, there is a lack of systematic T&E strategies throughout the ML-enabled system's lifecycle. This leads practitioners to resort to ad-hoc T&E practices, which can undermine user confidence in the reliability of ML-enabled software systems. New systematic testing approaches, adequacy measurements, and metrics are required to address the T&E challenges across all stages of the ML-enabled system lifecycle.

Aligned large language models (LLMs) demonstrate exceptional capabilities in task-solving, following instructions, and ensuring safety. However, the continual learning aspect of these aligned LLMs has been largely overlooked. Existing continual learning benchmarks lack sufficient challenge for leading aligned LLMs, owing to both their simplicity and the models' potential exposure during instruction tuning. In this paper, we introduce TRACE, a novel benchmark designed to evaluate continual learning in LLMs. TRACE consists of 8 distinct datasets spanning challenging tasks including domain-specific tasks, multilingual capabilities, code generation, and mathematical reasoning. All datasets are standardized into a unified format, allowing for effortless automatic evaluation of LLMs. Our experiments show that after training on TRACE, aligned LLMs exhibit significant declines in both general ability and instruction-following capabilities. For example, the accuracy of llama2-chat 13B on gsm8k dataset declined precipitously from 28.8\% to 2\% after training on our datasets. This highlights the challenge of finding a suitable tradeoff between achieving performance on specific tasks while preserving the original prowess of LLMs. Empirical findings suggest that tasks inherently equipped with reasoning paths contribute significantly to preserving certain capabilities of LLMs against potential declines. Motivated by this, we introduce the Reasoning-augmented Continual Learning (RCL) approach. RCL integrates task-specific cues with meta-rationales, effectively reducing catastrophic forgetting in LLMs while expediting convergence on novel tasks.

LLMs have shown great capabilities in various tasks but also exhibited memorization of training data, thus raising tremendous privacy and copyright concerns. While prior work has studied memorization during pre-training, the exploration of memorization during fine-tuning is rather limited. Compared with pre-training, fine-tuning typically involves sensitive data and diverse objectives, thus may bring unique memorization behaviors and distinct privacy risks. In this work, we conduct the first comprehensive analysis to explore LMs' memorization during fine-tuning across tasks. Our studies with open-sourced and our own fine-tuned LMs across various tasks indicate that fine-tuned memorization presents a strong disparity among tasks. We provide an understanding of this task disparity via sparse coding theory and unveil a strong correlation between memorization and attention score distribution. By investigating its memorization behavior, multi-task fine-tuning paves a potential strategy to mitigate fine-tuned memorization.

In recent years, the remarkable progress of Machine Learning (ML) technologies within the domain of Artificial Intelligence (AI) systems has presented unprecedented opportunities for the aviation industry, paving the way for further advancements in automation, including the potential for single pilot or fully autonomous operation of large commercial airplanes. However, ML technology faces major incompatibilities with existing airborne certification standards, such as ML model traceability and explainability issues or the inadequacy of traditional coverage metrics. Certification of ML-based airborne systems using current standards is problematic due to these challenges. This paper presents a case study of an airborne system utilizing a Deep Neural Network (DNN) for airport sign detection and classification. Building upon our previous work, which demonstrates compliance with Design Assurance Level (DAL) D, we upgrade the system to meet the more stringent requirements of Design Assurance Level C. To achieve DAL C, we employ an established architectural mitigation technique involving two redundant and dissimilar Deep Neural Networks. The application of novel ML-specific data management techniques further enhances this approach. This work is intended to illustrate how the certification challenges of ML-based systems can be addressed for medium criticality airborne applications.

With leaps in machine learning techniques and their applicationon Earth observation challenges has unlocked unprecedented performance across the domain. While the further development of these methods was previously limited by the availability and volume of sensor data and computing resources, the lack of adequate reference data is now constituting new bottlenecks. Since creating such ground-truth information is an expensive and error-prone task, new ways must be devised to source reliable, high-quality reference data on large scales. As an example, we showcase E URO C ROPS, a reference dataset for crop type classification that aggregates and harmonizes administrative data surveyed in different countries with the goal of transnational interoperability.

Watermark algorithms for large language models (LLMs) have achieved extremely high accuracy in detecting text generated by LLMs. Such algorithms typically involve adding extra watermark logits to the LLM's logits at each generation step. However, prior algorithms face a trade-off between attack robustness and security robustness. This is because the watermark logits for a token are determined by a certain number of preceding tokens; a small number leads to low security robustness, while a large number results in insufficient attack robustness. In this work, we propose a semantic invariant watermarking method for LLMs that provides both attack robustness and security robustness. The watermark logits in our work are determined by the semantics of all preceding tokens. Specifically, we utilize another embedding LLM to generate semantic embeddings for all preceding tokens, and then these semantic embeddings are transformed into the watermark logits through our trained watermark model. Subsequent analyses and experiments demonstrated the attack robustness of our method in semantically invariant settings: synonym substitution and text paraphrasing settings. Finally, we also show that our watermark possesses adequate security robustness. As the quality of text generated by large language models (LLMs) continues to improve, it addresses a multitude of practical challenges on one hand, while simultaneously giving rise to a spectrum of new issues on the other. Therefore, the detection and labeling of machine-generated text have become extremely important. Text watermarking techniques for LLMs usually embed specific information during text generation to allow high-accuracy detection of LLM-generated text. The mainstream approach for embedding such information is to add extra watermark logits on top of the logits generated by the LLM. For example, Kirchenbauer et al. (2023) divide the vocabulary into red and green lists and increase the scores for the green tokens as the watermark logits.