The multi-domain fine-tuning of large language models (LLMs) confronts a notorious trade-off among abilities across domains. Existing studies attribute this trade-off to the conflicts between samples rooted in inherent semantics. Recent approaches attempt to mitigate these conflicts through the empirical investigation or heuristic strategies. However, without a fundamental understanding of interactions between samples, they yield only marginal improvements, while incurring substantial trial-and-error costs. To address this challenge, we move beyond empirical studies by modeling interactions between samples as their influence on each other's loss, estimated using gradients. Intriguingly, we find that these interactions **evolve throughout training** rather than being purely determined by inherent semantics. Building on this insight, we propose **EV**olving **I**nteraction-guided **C**urriculum (**EVIC**), which iteratively selects samples that positively influence the overall dataset for training. By dynamically adapting the training curriculum to prioritize samples that contribute the most to the model training, EVIC effectively mitigates conflicts and improves the sample efficiency. Extensive experiments on a mixed dataset covering coding, math, and general tasks with several model architectures show that EVIC significantly outperforms all baselines across diverse capabilities.

Visible light positioning (VLP) is a promising direction for indoor localization. VLP depends on Visible light communication (VLC) to receive location anchors sent by light bulbs. In order to decode high-frequency VLC signals, today's VLP systems require the receiver to equip either rolling shutter cameras or high-frequency light sensors, which bring considerable overhead or are even unavailable on many mobile devices. This paper introduces ALS-P, a lightweight VLP approach which only requires the commercially widely available ambient light sensor (ALS). ALS is conventionally not treated as a feasible VLC receiver as its sampling rate is far less than that of VLC signals. Our basic idea is to leverage the property of frequency aliasing. Through dynamically adjusting the sampling rate of the ALS sensor, the down-converted signals can be uniquely distinguished. To realize this idea, we propose novel designs to address challenges which stem from ALS hardware, high-order light aliasing, and environmental interference cancellation. Besides, ALS can also enable a lightweight VLC via changing LED frequencies. We implement the ALS-P on commercial LED bulbs and the ALS of existing smartphones. The evaluation shows that the ALSP can enable robust and efficient LED frequency decoding with at least 4LEDs bulbs in practical scenarios. As a result, ALS-P can enable VLC at a data rate of 5bit/s per each LED bulb and achieve a sub-meter level indoor localization accuracy.

The conventional high-speed Wi-Fi has recently become a contender for low-power Internet-of-Things (IoT) communications. OFDM continues its adoption in the new IoT Wi-Fi standard due to its spectrum efficiency that can support the demand of massive IoT connectivity. While the IoT Wi-Fi standard offers many new features to improve power and spectrum efficiency, the basic physical layer (PHY) structure of transceiver design still conforms to its conventional design rationale where access points (AP) and clients employ the same OFDM PHY. In this paper, we argue that current Wi-Fi PHY design does not take full advantage of the inherent asymmetry between AP and IoT. To fill the gap, we propose an asymmetric design where IoT devices transmit uplink packets using the lowest power while pushing all the decoding burdens to the AP side. Such a design utilizes the sufficient power and computational resources at AP to trade for the transmission (TX) power of IoT devices. The core technique enabling this asymmetric design is that the AP takes full power of its high clock rate to boost the decoding ability. We provide an implementation of our design and show that it can reduce the IoT's TX power by boosting the decoding capability at the receivers.

Since the mobile camera is often small in size and easy to conceal, existing anti-piracy solutions are inefficient in defeating mobile-camera-based piracy, thus it remains a serious threat to copyright of Intellectual property. This paper presents Rainbow, a low-cost lighting system which prevents mobile-camera-based piracy attacks on intellectual property in the physical world, e.g., art paintings. Through embedding invisible illuminance flickers and chromatic changes into the light, our system can significantly degrade the imaging quality of a camera while maintaining a good visual experience for human eyes. Extensive objective evaluations under different scenarios demonstrate that Rainbow is robust to different confounding factors and can effectively defeat piracy attacks on various mobile devices. Subjective tests on volunteers further evidence that our system can not only significantly pollute pirated photos but also be able to provide good lighting conditions.

On-body devices are an intrinsic part of the Internet-of-Things (IoT) vision to provide human-centric services. These on-body IoT devices are largely embedded devices that lack a sophisticated user interface to facilitate traditional pre-shared key-based security protocols. Motivated by this real-world security vulnerability, this paper proposes SecureTag, a system designed to add defense in depth against active attacks by integrating physical layer (PHY) information with upper-layer protocols. The underpinning of SecureTag is a signal processing technique that extracts the peculiar propagation characteristics of creeping waves to discern on-body devices. Upon overhearing a suspicious transmission, SecureTag initiates a PHY-based challenge-response protocol to mitigate attacks. We implement our system on different commercial off-the-shelf wearables and a smartphone. Extensive experiments are conducted in a lab, apartments, malls, and outdoor areas, involving 12 volunteer subjects of different age groups, to demonstrate the robustness of our system. Results show that our system can mitigate 96.13 % of active attack attempts while triggering false alarms on merely 5.64 % of legitimate traffic.

Securely pairing wearables with ano-ther device is the key to many promising applications. This paper presents Touch-And-Guard (TAG), a system that uses hand touch as an intuitive manner to establish a secure connection between a wristband wearable and the touched device. It generates secret bits from hand resonant properties, which are obtained using accelerometers and vibration motors. The extracted secret bits are used by both sides to authenticate each other and then communicate confidentially. The ubiquity of accelerometers and motors presents an immediate market for our system. We demonstrate the feasibility of our system using an experimental prototype and conduct experiments involving 12 participants with 1,440 trials. The results indicate that we can generate secret bits at a rate of 7.15 bit/s, which is 44 percent faster than conventional text input PIN authentication.

The ability to detect which wearables and smartphones are on the same body has the potential to support a wealth of applications, including user authentication, automatic data synchronization, and personalized profile loading. This paper brings this feature to commercial off-the-shelf (COTS) wearables and smartphones, by creating a virtual “on-body detection sensor” based on devices' inherent wireless capabilities. We investigate using the peculiar propagation characteristics of creeping waves to discern on-body wearables. To this end, we decompose signals into multiple independent components to exploit the variation features of creeping waves. We implement our system on COTS wearables and a smartphone. Extensive experiments are conducted in a lab, apartments, malls, and outdoor areas, involving 12 volunteer subjects of different age groups, to demonstrate the robustness of our system. Results show that our system can identify on-body devices at 92.3% average true positive rate and 5% average false positive rate.

Securely pairing wearables with another device is the key to many promising applications, such as mobile payment, sensitive data transfer and secure interactions with smart home devices. This paper presents Touch-And-Guard (TAG), a system that uses hand touch as an intuitive manner to establish a secure connection between a wristband wearable and the touched device. It generates secret bits from hand resonant properties, which are obtained using accelerometers and vibration motors. The extracted secret bits are used by both sides to authenticate each other and then communicate confidentially. The ubiquity of accelerometers and motors presents an immediate market for our system. We demonstrate the feasibility of our system using an experimental prototype and conduct experiments involving 12 participants with 1440 trials. The results indicate that we can generate secret bits at a rate of 7.84 bit/s, which is 58% faster than conventional text input PIN authentication. We also show that our system is resistant to acoustic eavesdroppers in proximity.

Owing to the proliferation of mobile devices and their corresponding app ecosystems, more and more people are accessing the internet via various mobile apps, which generates tremendous volume of mobile data. Despite the growing importance of these mobile apps, we have a rather sparse understanding of how they are accessed and what issues affect their usage patterns. To address this problem, we perform a comprehensive measurement on large-scale anonymized network data collected from a tier-1 cellular carrier in China. In this measurement, we characterize the usage pattern of mobile apps and exhibit how the mobility, geospatial properties and behaviours of subscribers affect their mobile app usage at a fine-grained level.

User identification is an essential problem for security protection and data privacy preservation of wearable devices. With proper user identification, wearable devices can adopt personalized settings for different users, automatically label the corresponding data to protect user privacy, and help prevent illegal user spoofing attacks. Current user identification solutions proposed for wearable devices either rely on dedicated devices with high cost or require user intervention which is not convenient. In this work, we leverage the bio-vibrometry to enable a novel user identification solution for wearable devices in small-scale scenarios, e.g., household scenario. Unlike existing user identification solutions, our system only uses the low-cost sensors that are already available for most wearable devices. The key idea is that, when human body is exposed to a vibration excitation, the vibration response reflects the physical characteristics of user, i.e., the mass, stiffness and damping. Meanwhile, due to users' biological diversity, such physical characteristics of different users are quite distinctive. Therefore, we can leverage the discrepancy in users' vibration responses as an identifier. Based on this idea, we propose VibID, which only uses a low-cost vibration motor and accelerometer to generate an unobtrusive vibration to users' arms and capture the corresponding responses. By examining the vibration patterns at different frequencies, VibID builds an ensemble machine learning model to recognize who is using the device. Extensive experiments are conducted on human subjects to demonstrate that our system is reliable in small- scale scenarios and robust to various confounding factors, e.g., arm position, muscle state, user mobility and wearing location. We also show that, in an uncontrolled scenario of 8 users, our system can still ensure a identification accuracy above 91%.

Forming secure pairing between wearable devices has become an important problem in many scenarios, such as mobile payments and private data transmission. This paper presents EMG-KEY, a system that can securely pair wearable devices by leveraging the electrical activity caused by human muscle contraction, that is, Electromyogram (EMG), to generate a secret key. Such a key can then be used by devices to authenticate each other's physical proximity and communicate confidentially. Extensive evaluation on 10 volunteers under different scenarios demonstrates that our system can achieve a competitive bit generation rate of 5.51 bit/s while maintaining a matching probability of 88.84%. Also, the evaluation results with the presence of adversaries demonstrate our system is secure to strong attackers who can eavesdrop on proximate wireless communication, capture and imitate legitimate pairing process with the help of camera.

Similar to TV bands, the guardband frequencies are not occupied therefore are whitespace that potentially allows additional communication activities. Considering the difference to TV whitespace, we propose independent communication for guardband whitespace. In this paper, we present the Pilotfish system which realizes independent communication and turns guardband whitespace into new communication channels. To address the big challenges of interference mitigation, we employ novel PHY design which includes specially customized FBMC and an Nulled Decoding technique to null the strong background signal in guardbands. We implemented Pilotfish using software radio system. Empirical evaluation results validate the Pilotfish design in both PHY and MAC.

Nowadays, cardiovascular disease (CVD) has become a disease of the majority. As an important instrument for diagnosing CVD, electrocardiography (ECG) is used to extract useful information about the functioning status of the heart. In the domain of ECG analysis, cluster analysis is a commonly applied approach to gain an overview of the data, detect outliers or pre-process before further analysis. In recent years, to provide better medical care for CVD patients, the cardiac telehealth system has been widely used. However, the extremely large volume and high update rate of data in the telehealth system has made cluster analysis challenging work. In this paper, we design and implement a novel parallel system for clustering massive ECG stream data based on the MapReduce framework. In our approach, a global optimum of clustering is achieved by merging and splitting clusters dynamically. Meanwhile, a good performance is gained by distributing computation over multiple computing nodes. According to the evaluation, our system not only provides good clustering results but also has an excellent performance on multiple computing nodes.