Real-time health monitoring systems generate massive biometric sensor data, placing substantial demands on memory, computation, and power resources. Additionally, the sensitive nature of such data raises critical privacy concerns related to attributes like gender, age, and body mass index. To address these challenges, this work proposes FAxC—a novel feature-driven approximation framework that performs multi-dimensional data reduction by leveraging the biometric features of sensor signals. Unlike prior approximation techniques that operate on isolated signals or uniformly sample data, FAxC intelligently selects and masks segments to preserve human activity recognition performance while enhancing user privacy. Our case study shows that, compared to existing privacy-aware approaches, FAxC reduces the disparity in gender-distinguishable biometric features in human activity recognition, decreasing Z-direction peak acceleration from 56% to 22% and stride root mean square from 31% to 3%. Our FAxC enhances the privacy-preserving rate by up to $5.6 \times$ over the baseline and outperforms existing methods by a factor ranging from $1.2 \times$ to $5.4 \times$. The proposed method has also been validated on the TAME Pain dataset for a voice-based pain level detection system. FAxC reduces the risk of gender leakage by 50% compared to using raw data, while maintaining pain level detection accuracy.

In-Sensor Computing (ISC) systems emerge as a promising alternative to save energy on massive data transmission, analog-to-digital conversion, and ineffective processing. While the new paradigm shift of ISC systems gains increasing attention, the highly compacted systems could incur new challenges from a hardware security perspective. This work conducts a literature review to highlight the research trend of this topic and then performs comprehensive analyses on the root of security challenges. To the best of our knowledge, this is the first work that compares the security challenges of traditional sensor-involved computing systems and emerging ISC systems. We conduct a comprehensive analysis of ISC's design phases to identify significant vulnerabilities and attack surfaces. Furthermore, new attack scenarios are predicted for board-, chip-, and device-level ISC systems. Three proof-of-concept demos are provided to reveal the consequences of the attacks for three different levels. Our findings emphasize the urgent necessity for sophisticated defense mechanisms and inspire researchers to work on new countermeasure designs against unique hardware security threats in ISC systems.

FPGAs are being challenged by various security threats, including reverse engineering attacks, hardware tampering, and side-channel analysis attacks. Although the existing static obfuscation methods can protect FPGA systems from IP piracy and hardware tampering, limited work is available to improve the attack resilience of obfuscation modules. As hardware Trojans are one of the most significant hardware tampering attacks on FPGAs, this work aims for the specific hardware Trojan that attempts to nullify design obfuscation. To address this need, we leverage the advanced function of FPGA CAD tools to propose a Dynamic Partial Reconfiguration enabled Design Obfuscation (DPReDO) method. Our method partially modifies the FPGA bitstream at runtime to remove the sabotaged obfuscation variant, thus offering enhanced attack resilience against hardware Trojans. Experimental results based on ISCAS and ITC-99 benchmark circuits show that the DPReDO method reduces the Trojan hit rate by up to 80% over existing static obfuscation with less than 3% hardware overhead. To test the practical feasibility of the proposed countermeasure, we further apply DPReDO to an FPGA-accelerated computation engine for a financial application. Compared to static obfuscation, the proposed DPReDO only incurs 2.6% and 1.2% more FPGA LUTs and slices, respectively.

Real-time health monitoring systems generate a large volume of sensing data, requiring tremendous processing time and storage space. Orthogonal to existing approximate computing mechanisms, this work proposes a Feature-Driven Approximation (FDApx) method to address the pressing need for fast data processing and a limited storage budget in wearable health monitoring devices. The proposed FDApx method reverses the features interested in high-level applications to derive approximation thresholds to retain feature-critical information, rather than aimlessly storing and transmitting all raw data. Case studies in an insole sensing system for fall risk assessment show that FDApx can reduce the data size by up to 87% over raw data and up to 85% over 2-bit precision reduction-based approximation. The approximation from FDApx only results in up to a 2% deviation in swing time; in contrast, the approximation based on precision reduction causes a 30% deviation in the same gait feature1.

Approximate computing techniques that trade accuracy for better computing performance and energy efficiency have been widely used in many computation-intensive applications. As reported in the recent literature, approximate computing systems are prone to stealthy attacks that exploit approximation mechanisms to disguise malicious behaviors in normal operations. The analysis performed in this work indicates that the primary outputs of applications with approximate components are not the best location to detect the presence of attacks. This work proposes an intermediate node evaluation-based attack detection (INEAD) method to distinguish whether the cause of inaccuracy in applications is due to approximation or attacks. The attack detection rate of the proposed method was examined in two applications: 1) an approximate fast Fourier transform (FFT) and 2) an artificial neural network (ANN) using approximate arithmetic modules. The case study of an approximate FFT shows that the INEAD method improves the attack detection rate by 61.6% and reduces the false positive rate by up to 91% over the existing work that detects attacks at the primary output stage. The case study of an approximate ANN shows that the INEAD method outperforms the baseline by 93.6% in terms of attack detection rate.