Variances in ad impression outcomes across demographic groups are increasingly considered to be potentially indicative of algorithmic bias in personalized ads systems. While there are many definitions of fairness that could be applicable in the context of personalized systems, we present a framework which we call the Variance Reduction System (VRS) for achieving more equitable outcomes in Meta's ads systems. VRS seeks to achieve a distribution of impressions with respect to selected protected class (PC) attributes that more closely aligns the demographics of an ad's eligible audience (a function of advertiser targeting criteria) with the audience who sees that ad, in a privacy-preserving manner. We first define metrics to quantify fairness gaps in terms of ad impression variances with respect to PC attributes including gender and estimated race. We then present the VRS for re-ranking ads in an impression variance-aware manner. We evaluate VRS via extensive simulations over different parameter choices and study the effect of the VRS on the chosen fairness metric. We finally present online A/B testing results from applying VRS to Meta's ads systems, concluding with a discussion of future work. We have deployed the VRS to all users in the US for housing ads, resulting in significant improvement in our fairness metric. VRS is the first large-scale deployed framework for pursuing fairness for multiple PC attributes in online advertising.

With the coming era of cloud technology, cloud storage is an emerging technology to store massive digital images, which provides steganography a new fashion to embed secret information into massive images. Specifically, a resourceful steganographer could embed a set of secret information into multiple images adaptively, and share these images in cloud storage with the receiver, instead of traditional single image steganography. Nevertheless, it is still an open issue how to allocate embedding payload among a sequence of images for security performance enhancement. This paper formulates adaptive payload distribution in multiple images steganography based on image texture features and provides the theoretical security analysis from the steganalyst's point of view. Two payload distribution strategies based on image texture complexity and distortion distribution are designed and discussed respectively. The proposed strategies can be employed together with these state-of-the-art single image steganographic algorithms. The comparisons of the security performance against the modern universal pooled steganalysis are given. Furthermore, this paper compares the per image detectability of these multiple images steganographic schemes against the modern single image steganalyzer. Extensive experimental results show that the proposed payload distribution strategies could obtain better security performance.

Recent studies have shown that physiological signals such as heart beat and breathing can be remotely captured from human faces using a regular color camera under ambient light. This technology, referred to as remote photoplethysmography (rPPG), can be used to collect the physiological status of users who are in front of a camera, which may raise privacy concerns. To avoid the privacy abuse of the rPPG technology, this paper develops PulseEdit, a novel and efficient algorithm that can edit the physiological signals in facial videos without affecting visual appearance and thus protect the user’s physiological signal from disclosure. PulseEdit can either remove the trace of the physiological signal in a video or transform the video to contain a target physiological signal chosen by a user. Experimental results show that PulseEdit can effectively edit physiological signals in facial videos and prevent heart rate measurement based on rPPG. It is possible to utilize PulseEdit in adversarial scenarios against rPPG-based visual security algorithms. We present analyses on the performance of PulseEdit against rPPG-based liveness detection and rPPG-based deepfake detection, and demonstrate its ability to circumvent these visual security algorithms and its important role in supporting the design of attack-resilient systems.

In the field of information forensics, many emerging problems involve a critical step that estimates and tracks weak frequency components in noisy signals. It is often challenging for the prior art of frequency tracking to i) achieve a high accuracy under noisy conditions, ii) detect and track multiple frequency components efficiently, or iii) strike a good trade-off of the processing delay versus the resilience and the accuracy of tracking. To address these issues, we propose Adaptive Multi-Trace Carving (AMTC), a unified approach for detecting and tracking one or more subtle frequency components under very low signal-to-noise ratio (SNR) conditions and in near real time. AMTC takes as input a time-frequency representation of the system's preprocessing results (such as the spectrogram), and identifies frequency components through iterative dynamic programming and adaptive trace compensation. The proposed algorithm considers relatively high energy traces sustaining over a certain duration as an indicator of the presence of frequency/oscillation components of interest and track their time-varying trend. Extensive experiments using both synthetic data and real-world forensic data of power signatures and physiological monitoring reveal that the proposed method outperforms representative prior art under low SNR conditions, and can be implemented in near real-time settings. The proposed AMTC algorithm can empower the development of new information forensic technologies that harness very small signals.

This paper introduces an amplitude and frequency modulation (AM-FM) model to characterize the photoplethysmography (PPG) signal. The model indicates that the PPG signal spectrum contains one dominant frequency component - the heart rate (HR), which is guarded by two weaker frequency components on both sides; the distance from the dominant component to the guard components represents the respiratory rate (RR). Based on this model, an efficient algorithm is proposed to estimate both HR and RR by searching for the dominant frequency component and two guard components. The proposed method is performed in the frequency domain to estimate RR, which is more robust to additive noise than the prior art based on temporal features. Experiments were conducted on two types of PPG signals collected with a contact sensor (an oximeter) and a contactless visible imaging sensor (a color camera), respectively. The PPG signal from the contactless sensor is much noisier than the signal from the contact sensor. The experimental results demonstrate the effectiveness of the proposed algorithm, including under relatively noisy scenarios.

Entropy coding is a lossless data compression technique that is widely applied in video codecs to encode syntax elements into bitstreams. Efficient entropy coding requires accurate prediction of the probability distribution of the encoded symbols. In AV1, multi-symbol arithmetic coding is adopted. The symbol probability is derived with handcrafted context models and lookup tables that store the predicted probabilities corresponding to different entropy contexts. The lookup table based scheme has some fundamental deficiencies. The entropy context features have to be discrete so that they can be used to index the lookup tables. To reduce the size of the lookup table, the number of contexts cannot be very large. Moreover, the probability distributions stored in the lookup tables are maintained separately without taking their correlations into consideration. In this paper, we propose a machine learning based scheme that achieves more accurate symbol probability prediction for entropy coding. The proposed approach is implemented in AV1 for the entropy coding of intra prediction modes. Experimental results demonstrate that it can improve the efficiency of entropy coding significantly.

Effective machine learning models can automatically learn useful information from a large quantity of data and provide decisions in a high accuracy. These models may, however, lead to unfair predictions in certain sense among the population groups of interest, where the grouping is based on such sensitive attributes as race and gender. Various fairness definitions, such as demographic parity and equalized odds, were proposed in prior art to ensure that decisions guided by the machine learning models are equitable. Unfortunately, the "fair" model trained with these fairness definitions is threshold sensitive, i.e., the condition of fairness may no longer hold true when tuning the decision threshold. This paper introduces the notion of threshold invariant fairness, which enforces equitable performances across different groups independent of the decision threshold. To achieve this goal, this paper proposes to equalize the risk distributions among the groups via two approximation methods. Experimental results demonstrate that the proposed methodology is effective to alleviate the threshold sensitivity in machine learning models designed to achieve fairness.