Search for: All records

Differentially Private Stochastic Gradient Descent (DP-SGD) has become a widely used technique for safeguarding sensitive information in deep learning applications. Unfortunately, DP-SGD’s per-sample gradient clipping and uniform noise addition during training can significantly degrade model utility and fairness. We observe that the latest DP-SGD-Global-Adapt’s average gradient norm is the same throughout the training. Even when it is integrated with the existing linear decay noise multiplier, it has little or no advantage. Moreover, we notice that its upper clipping threshold increases exponentially towards the end of training, potentially impacting the model’s convergence. Other algorithms, DP-PSAC, Auto-S, DP-SGD-Global, and DP-F, have utility and fairness that are similar to or worse than DP-SGD, as demonstrated in experiments. To overcome these problems and improve utility and fairness, we developed the DP-SGD-Global-Adapt-V2-S. It has a step-decay noise multiplier and an upper clipping threshold that is also decayed step-wise. DP-SGD-Global-Adapt-V2-S with a privacy budget of 1 improves accuracy by 0.9795%, 0.6786%, and 4.0130% in MNIST, CIFAR10, and CIFAR100, respectively. It also reduces the privacy cost gap by 89.8332% and 60.5541% in unbalanced MNIST and Thinwall datasets, respectively. Finally, we develop mathematical expressions to compute the privacy budget using truncated concentrated differential privacy (tCDP) for DP-SGD-Global-Adapt-V2-T and DP-SGD-Global-Adapt-V2-S. 

Recent advancements in person recognition have raised concerns about identity privacy leaks. Gait recognition through millimeter-wave radar provides a privacy-centric method. However, it is challenged by lower accuracy due to the sparse data these sensors capture. We are the first to investigate a cross-modal method, IdentityKD, to enhance gait-based person recognition with the assistance of facial data. IdentityKD involves a training process using both gait and facial data, while the inference stage is conducted exclusively with gait data. To effectively transfer facial knowledge to the gait model, we create a composite feature representation using contrastive learning. This method integrates facial and gait features into a unified embedding that captures the unique identityspecific information from both modalities. We employ two distinct contrastive learning losses. One minimizes the distance between embeddings of data pairs from the same person, enhancing intraclass compactness, while the other maximizes the distance between embeddings of data pairs from different individuals, improving inter-class separability. Additionally, we use an identity-wise distillation strategy, which tailors the training process for each individual, ensuring that the model learns to distinguish between different identities more effectively. Our experiments on a dataset of 36 subjects, each providing over 5000 face-gait pairs, demonstrate that IdentityKD improves identity recognition accuracy by 6.5% compared to baseline methods. 

In various scenarios from system login to writing emails, documents, and forms, keyboard inputs carry alluring data such as passwords, addresses, and IDs. Due to commonly existing non-alphabetic inputs, punctuation, and typos, users' natural inputs rarely contain only constrained, purely alphabetic keys/words. This work studies how to reveal unconstrained keyboard inputs using auditory interfaces. Audio interfaces are not intended to have the capability of light sensors such as cameras to identify compactly located keys. Our analysis shows that effectively distinguishing the keys can require a fine localization precision level of keystroke sounds close to the range of microseconds. This work (1) explores the limits of audio interfaces to distinguish keystrokes, (2) proposes a μs-level customized signal processing and analysis-based keystroke tracking approach that takes into account the mechanical physics and imperfect measuring of keystroke sounds, (3) develops the first acoustic side-channel attack study on unconstrained keyboard inputs that are not purely alphabetic keys/words and do not necessarily follow known sequences in a given dictionary or training dataset, and (4) reveals the threats of non-line-of-sight keystroke sound tracking. Our results indicate that, without relying on vision sensors, attacks using limited-resolution audio interfaces can reveal unconstrained inputs from the keyboard with a fairly sharp and bendable "auditory eyesight."