In the era of costly pre-training of large language models, ensuring the intellectual property rights of model owners, and insuring that said models are responsibly deployed, is becoming increasingly important. To this end, we propose model watermarking via passthrough layers, which are added to existing pre-trained networks and trained using a self-supervised loss such that the model produces high-entropy output when prompted with a unique private key, and acts normally otherwise. Unlike existing model watermarking methods, our method is fully task-agnostic, and can be applied to both classification and sequence-to-sequence tasks without requiring advanced access to downstream fine-tuning datasets. We evaluate the proposed passthrough layers on a wide range of downstream tasks, and show experimentally our watermarking method achieves a near-perfect watermark extraction accuracy and false-positive rate in most cases without damaging original model performance. Additionally, we show our method is robust to both downstream fine-tuning, fine-pruning, and layer removal attacks, and can be trained in a fraction of the time required to train the original model. Code is available in the paper.

Convolutional Neural Networks (CNNs) are used mainly for image classification because of their high accuracy and fast performance. Due to their complexity, their functionality is like a black box to the human user. Hence, this black box functionality may classify an image based on the wrong features, which can lead to severe consequences in critical applications, such as disease detection. Therefore, explainer methods provide the users with the reasoning behind each classification. However, selecting well-matching pairs of CNN models and explainers is challenging.This paper proposes a framework for automated explainability evaluation in CNN models, which follows a quantitative approach for assessing the model and explainer pairs. The proposed framework supersedes the previous attempts towards an explainability evaluation framework by replacing the time-consuming qualitative measure with a unified quantitative metric. This quantitative approach allows the users to assess several models and explainer pairs and compare the results based on two aspects on the one hand, selecting the most prominent model/explainer pair or, on the other hand, making compromises for better real-time performance. The framework is applied to a defect detection problem for printed circuit board assembly(PCBA) automatic optical inspection (AOI). The results are analyzed for thirteen built-in CNN models and ten built-in explainer methods. The results demonstrate the superiority of the "CoAtNet0" and "Grad-Cam selfmatch" pair with the 55.38% Explanation Score (ES) metric. The authors also provide a discussion on other criteria in the selection of a prominent pair using the proposed novel FACEE framework.

Developing wearable platforms for unconstrained monitoring of limb movements has been an active recent topic of research due to potential applications such as clinical and athletic performance evaluation. However, practicality of these platforms might be affected by the dynamic and complexity of movements as well as characteristics of the surrounding environment. This paper addresses such issues by proposing a novel method for obtaining kinematic information of joints using a custom-designed wearable platform. The proposed method uses data from two gyroscopes and an array of textile stretch sensors to accurately track three-dimensional movements, including extension, flexion, and rotation, of a joint. More specifically, gyroscopes provide angular velocity data of two sides of a joint, while their relative orientation is estimated by a machine learning algorithm. An Unscented Kalman Filter (UKF) algorithm is applied to directly fuse angular velocity/relative orientation data and estimate the kinematic orientation of the joint. Experimental evaluations were carried out using data from 10 volunteers performing a series of predefined as well as unconstrained random three-dimensional trunk movements. Results show that the proposed sensor setup and the UKF-based data fusion algorithm can accurately estimate the orientation of the trunk relative to pelvis with an average error of less than 1.72 degrees in predefined movements and a comparable accuracy of 3.00 degrees in random movements. Moreover, the proposed platform is easy to setup, does not restrict body motion, and is not affected by environmental disturbances. This study is a further step towards developing user-friendly wearable sensor systems than can be readily used in indoor and outdoor settings without requiring bulky equipment or a tedious calibration phase.