Accurate swimming performance monitoring has traditionally relied on wearable sensors, which can disrupt natural technique and are impractical in competitive settings. In this paper, we present a fully vision-based system for automatic swimmer analysis using overhead drone footage, removing the need for any wearable device or underwater equipment. By fine-tuning pose estimation models for aerial aquatic conditions, our approach robustly extracts full-body swimmer skeletons even under challenging scenarios such as splashes and partial occlusions. From these poses, we classify swimming strokes, compute instantaneous speed, estimate lap times, and count individual strokes. Unlike existing methods, our system provides scalable, unobtrusive, and infrastructure-free tracking. Evaluated on real-world drone-captured swimming competition data, our method achieves a median speed estimation error below 4% (under 0.05 m/s), a median lap time error of just 0.03s, and stroke count errors typically under one stroke per lap.

With the aggressive growth in AI model complexity to achieve higher performance, operating them on mobile platforms becomes more and more challenging. This issue is even more prominent for time-series data, commonly dealt with in mobile/IoT computing scenarios, given their inherent issues such as label imbalance, user and sensor diversity, and out-of-distribution inference data. In this work, we investigate into the efficacy of a k-nearest neighbor classifier enhanced with a lossless compressor gzip, introducing novel sequence tokenization algorithms that show superior performance compared to traditional machine/deep learning classifiers. Our evaluation across three diverse real-world applications with distinct datasets emphasizes the generalization potential of our approach in real-world scenarios, especially in situations with few training samples.

The prevalence of counterfeit infant formulas worldwide poses serious threats to infant health and safety, a concern highlighted by the notorious Melamine Milk Scandal that affected hundreds of thousands of children. The primary challenge in detecting counterfeit formulas lies in their sophisticated adulteration and substitution techniques. Such detection is feasible only in laboratory settings, making it nearly impossible for average consumers to test the formula before feeding their infants. To address this problem, we propose PowDew, a novel and practical system for detecting counterfeit infant formula that utilizes only a commodity smartphone. PowDew operates by capturing and analyzing the interaction of a water droplet with the powdered formula, focusing on the droplet motion, namely its spreading and penetration. Our insight is that the droplet motions are governed by powder-specific properties such as wettability and porosity. PowDew analyzes the subtle differences in droplet motions, and infers the formula's authenticity. To demonstrate PowDew's effectiveness, we implement PowDew and conduct comprehensive real-world experiments under varying conditions with different brands of powdered infant formula and adulterants. Our experiments result in a total of 12,000 minutes of video recordings of the droplet motions on various infant formulas, including authentic and altered. Our experiments demonstrate that PowDew yields an overall detection accuracy of up to 96.1%.

The rise of counterfeit powdered food products, exemplified by notorious incidents such as the Melamine Milk Scandal, poses significant risks to consumers. The primary challenge in identifying these counterfeit products comes from their intricate adulteration and substitution techniques. Currently, such identification methods are only viable in laboratory settings, making average consumers nearly impossible to authenticate their products. To address this limitation, we propose PowDew, a novel system that employs a smartphone to detect counterfeit powdered food products. PowDew utilizes the powder's physical property, namely droplet motion, as a basis for verification. Through real-world experiments, PowDew demonstrate a practicality with achieving an overall detection accuracy of up to 96.1%.

While federated learning leverages distributed client resources, it faces challenges due to heterogeneous client capabilities. This necessitates allocating models suited to clients' resources and careful parameter aggregation to accommodate this heterogeneity. We propose HypeMeFed, a novel federated learning framework for supporting client heterogeneity by combining a multi-exit network architecture with hypernetwork-based model weight generation. This approach aligns the feature spaces of heterogeneous model layers and resolves per-layer information disparity during weight aggregation. To practically realize HypeMeFed, we also propose a low-rank factorization approach to minimize computation and memory overhead associated with hypernetworks. Our evaluations on a real-world heterogeneous device testbed indicate that HypeMeFed enhances accuracy by 5.12% over FedAvg, reduces the hypernetwork memory requirements by 98.22%, and accelerates its operations by 1.86X compared to a naive hypernetwork approach. These results demonstrate HypeMeFed's effectiveness in leveraging and engaging heterogeneous clients for federated learning.

In this demonstration, we introduce the vulnerability of indices in unpooling autoencoders. We show that this small factor can be maliciously exploited by performing man-in-the-middle attacks to eavesdrop on the victim's data, resulting in reconstruction and adversarial attacks. Such attacks especially make systems that integrate collaborative inference operations vulnerable. This demo presentation will empirically show the feasibility of index-based attacks by launching reconstruction and adversarial attacks on embedded/mobile computing platforms.

Radio-frequency catheter ablation (RCA) is an established treatment for ventricular tachycardia (VT). A key feature of the RCA procedure is the need for a mapping approach that facilitates the identification of the target ablation site. In this study, we investigate the effect of the location of the reference potential and spatial anatomical constraints on the accuracy of an algorithm to identify the target site for ablation therapy of VT. This algorithm involves processing body surface potentials using the single equivalent moving dipole (SEMD) model embedded in an infinite homogeneous volume conductor to model cardiac electrical activity. We employed a swine animal model and an electrode array of nine electrodes that was sutured on the epicardial surface of the right ventricle. We identified two potential reference electrode locations: at an electrode most far away from the heart (R1) and at the average of all 64 body surface electrode potentials (R2). Also, we developed three spatial "constraining" schemes of the algorithm used to obtain the SEMD location: one that does not impose any constraint on the inverse solution (S1), one that constrains the solution into a volume that corresponds to the heart (S2), and one that constrains the solution into a volume that corresponds to the body surface (S3). We have found that R2S1 is the most accurate approach (p < 0.05 versus R1S1 at earliest activation time-EAT) for localizing epicardial electrical sources of known locations in vivo. Although the homogeneous volume conductor introduces systematic error in the estimated compared to the true dipole location, we have observed that the overall error of the estimated interelectrode distance compared to the true one was 0.4 ± 0.4 cm and 0.4 ± 0.1 cm for the R1S1 and R2S1 combinations, respectively, at the EAT (p = N.S.) and 1.0 ± 0.6 and 0.5 ± 0.4 cm, respectively, at the pacing spike time (PST, ). In conclusion, our algorithm to estimate the SEMD parameters from body surface potentials can potentially be a useful method to rapidly and accurately guide the catheter tip to the target site during a RCA procedure without the need for spatial anatomical information obtained by conventional imaging modalities.