Large Language Models (LLMs) are powerful tools used in several applications such as conversational AI, and code generation. However, significant robustness concerns arise with LLMs in production, such as hallucinations, prompt injection attacks, harmful content generation, and challenges in maintaining accurate domain-specific content moderation. Guardrails aim to mitigate these challenges by aligning LLM outputs with desired behaviors without modifying the underlying models. Nvidia NeMo Guardrails, for instance, rely on specifying acceptable/unacceptable behaviours. However, it is challenging to predict and address potential issues of LLMs in advance to create these guardrails. Also, manual updates from software engineers are often required to maintain and refine these guardrails. We introduce LLM-Guards, specialised machine learning (ML) models trained to function as protective guards. Additionally, we present an automation pipeline for training and continual fine-tuning of these guards using reinforcement learning from human feedback (RLHF). We evaluated several small LLMs, including Llama-3, Mistral, and Gemma, as LLM-Guards for challenges such as moderation and detecting off-topic queries, and compared their performance against NeMo Guardrails. The proposed Llama-3 LLM-Guard outperformed NeMo Guardrails in detecting offtopic queries, achieving an accuracy of 98.7% compared to 81%. Furthermore, the LLM-Guard detected 97.86% of harmful queries” surpassing NeMo Guardrails by 19.86%.

Machine learning (ML), especially with the emergence of large language models (LLMs), has significantly transformed various industries. However, the transition from ML model prototyping to production use within software systems presents several challenges. These challenges primarily revolve around ensuring safety, security, and transparency, subsequently influencing the overall robustness and trustworthiness of ML models. In this paper, we introduce ML-On-Rails, a protocol designed to safeguard ML models, establish a well-defined endpoint interface for different ML tasks, and clear communication between ML providers and ML consumers (software engineers). ML-On-Rails enhances the robustness of ML models via incorporating detection capabilities to identify unique challenges specific to production ML. We evaluated the ML-On-Rails protocol through a real-world case study of the MoveReminder application. Through this evaluation, we emphasize the importance of safeguarding ML models in production.

Machine learning (ML) models have become essential components in software systems across several domains, such as autonomous driving, healthcare, and finance. The robustness of these ML models is crucial for maintaining the software systems performance and reliability. A significant challenge arises when these systems encounter out-of-distribution (OOD) data, examples that differ from the training data distribution. OOD data can cause a degradation of the software systems performance. Therefore, an effective OOD detection mechanism is essential for maintaining software system performance and robustness. Such a mechanism should identify and reject OOD inputs and alert software engineers. Current OOD detection methods rely on hyperparameters tuned with in-distribution and OOD data. However, defining the OOD data that the system will encounter in production is often infeasible. Further, the performance of these methods degrades with OOD data that has similar characteristics to the in-distribution data. In this paper, we propose a novel OOD detection method using the Gini coefficient. Our method does not require prior knowledge of OOD data or hyperparameter tuning. On common benchmark datasets, we show that our method outperforms the existing maximum softmax probability (MSP) baseline. For a model trained on the MNIST dataset, we improve the OOD detection rate by 4% on the CIFAR10 dataset and by more than 50% for the EMNIST dataset.

The advances in machine learning (ML) have stimulated the integration of their capabilities into software systems. However, there is a tangible gap between software engineering and machine learning practices, that is delaying the progress of intelligent services development. Software organisations are devoting effort to adjust the software engineering processes and practices to facilitate the integration of machine learning models. Machine learning researchers as well are focusing on improving the interpretability of machine learning models to support overall system robustness. Our research focuses on bridging this gap through a methodology that evaluates the robustness of machine learning-enabled software engineering systems. In particular, this methodology will automate the evaluation of the robustness properties of software systems against dataset shift problems in ML. It will also feature a notification mechanism that facilitates the debugging of ML components.

Human posture estimation is a fundamental challenge in computer vision research. This is a task that has received substantial interest due to the importance of evaluating the human performance in several disciplines. The ultimate goal for the vision-based pose estimation task is the markerless accurate prediction of necessary postural information. This paper proposes a single shot deep human posture detection and estimation network. The proposed SSDPose architecture increments standard object detection networks to feature posture estimation. SSDPose is an end-to-end trainable model that detects and estimates the body posture from a single image. Further, our network has been trained to predict joint angles which are essential information for several domains such as biomechanic and ergonomic posture analysis. The reference joint angles have been generated using motion capture sequences and a novel inverse kinematics method. Experimental results demonstrate that SSDPose effectively detects and estimates the posture by achieving person mean average precision (mAP) of 98.2%, an average joint angles MAE of 3.16 ± 1.23 deg and an RMSE of 4.22 ± 1.73 deg at up to 30 FPS.