Child Face Age-Progression via Deep Feature Aging

Given a gallery of face images of missing children, state-of-the-art face recognition systems fall short in identifying a child (probe) recovered at a later age. We propose a feature aging module that can age-progress deep face features output by a face matcher. In addition, the feature aging module guides age-progression in the image space such that synthesized aged faces can be utilized to enhance longitudinal face recognition performance of any face matcher without requiring any explicit training. For time lapses larger than 10 years (the missing child is found after 10 or more years), the proposed age-progression module improves the closed-set identification accuracy of FaceNet from 16.53% to 21.44% and CosFace from 60.72% to 66.12% on a child celebrity dataset, namely ITWCC. The proposed method also outperforms state-of-the-art approaches with a rank-1 identification rate of 95.91%, compared to 94.91%, on a public aging dataset, FG-NET, and 99.58%, compared to 99.50%, on CACD-VS. These results suggest that aging face features enhances the ability to identify young children who are possible victims of child trafficking or abduction.

Generalizing Face Representation with Unlabeled Data

In recent years, significant progress has been made in face recognition due to the availability of large-scale labeled face datasets. However, since the faces in these datasets usually contain limited degree and types of variation, the models trained on them generalize poorly to more realistic unconstrained face datasets. While collecting labeled faces with larger variations could be helpful, it is practically infeasible due to privacy and labor cost. In comparison, it is easier to acquire a large number of unlabeled faces from different domains which would better represent the testing scenarios in real-world problems. We present an approach to use such unlabeled faces to learn generalizable face representations, which can be viewed as an unsupervised domain generalization framework. Experimental results on unconstrained datasets show that a small amount of unlabeled data with sufficient diversity can (i) lead to an appreciable gain in recognition performance and (ii) outperform the supervised baseline when combined with less than half of the labeled data. Compared with the state-of-the-art face recognition methods, our method further improves their performance on challenging benchmarks, such as IJB-B, IJB-C and IJB-S.

Towards Universal Representation Learning for Deep Face Recognition

Recognizing wild faces is extremely hard as they appear with all kinds of variations. Traditional methods either train with specifically annotated variation data from target domains, or by introducing unlabeled target variation data to adapt from the training data. Instead, we propose a universal representation learning framework that can deal with larger variation unseen in the given training data without leveraging target domain knowledge. We firstly synthesize training data alongside some semantically meaningful variations, such as low resolution, occlusion and head pose. However, directly feeding the augmented data for training will not converge well as the newly introduced samples are mostly hard examples. We propose to split the feature embedding into multiple sub-embeddings, and associate different confidence values for each sub-embedding to smooth the training procedure. The sub-embeddings are further decorrelated by regularizing variation classification loss and variation adversarial loss on different partitions of them. Experiments show that our method achieves top performance on general face recognition datasets such as LFW and MegaFace, while significantly better on extreme benchmarks such as TinyFace and IJB-S.

Fingerprint Spoof Detection: Temporal Analysis of Image Sequence

We utilize the dynamics involved in the imaging of a fingerprint on a touch-based fingerprint reader, such as perspiration, changes in skin color (blanching), and skin distortion, to differentiate real fingers from spoof (fake) fingers. Specifically, we utilize a deep learning-based architecture (CNN-LSTM) trained end-to-end using sequences of minutiae-centered local patches extracted from ten color frames captured on a COTS fingerprint reader. A time-distributed CNN (MobileNet-v1) extracts spatial features from each local patch, while a bi-directional LSTM layer learns the temporal relationship between the patches in the sequence. Experimental results on a database of 26,650 live frames from 685 subjects (1,333 unique fingers), and 32,910 spoof frames of 7 spoof materials (with 14 variants) shows the superiority of the proposed approach in both known-material and cross-material (generalization) scenarios. For instance, the proposed approach improves the state-of-the-art cross-material performance from TDR of 81.65% to 86.20% @ FDR = 0.2%.

Fingerprint Synthesis: Search with 100 Million Prints

Evaluation of large-scale fingerprint search algorithms has been limited due to lack of publicly available datasets. A solution to this problem is to synthesize a dataset of fingerprints with characteristics similar to those of real fingerprints. We propose a Generative Adversarial Network (GAN) to synthesize a fingerprint dataset consisting of 100 million fingerprint images. In comparison to published methods, our approach incorporates an identity loss which guides the generator to synthesize a diverse set of fingerprints corresponding to more distinct identities. To demonstrate that the characteristics of our synthesized fingerprints are similar to those of real fingerprints, we show that (i) the NFIQ quality value distribution of the synthetic fingerprints follows the corresponding distribution of real fingerprints and (ii) the synthetic fingerprints are more distinct than existing synthetic fingerprints (and more closely align with the distinctiveness of real fingerprints). We use our synthesis algorithm to generate 100 million fingerprint images in 17.5 hours on 100 Tesla K80 GPUs when executed in parallel. Finally, we report for the first time in open literature, search accuracy (DeepPrint rank-1 accuracy of 91.4%) against a gallery of 100 million fingerprint images (using 2,000 NIST SD4 rolled prints as the queries).

Universal Material Translator: Towards Spoof Fingerprint Generalization

Spoof detectors are classifiers that are trained to distinguish spoof fingerprints from bonafide ones. However, state of the art spoof detectors do not generalize well on unseen spoof materials. This study proposes a style transfer based augmentation wrapper that can be used on any existing spoof detector and can dynamically improve the robustness of the spoof detection system on spoof materials for which we have very low data. Our method is an approach for synthesizing new spoof images from a few spoof examples that transfers the style or material properties of the spoof examples to the content of bonafide fingerprints to generate a larger number of examples to train the classifier on. We demonstrate the effectiveness of our approach on materials in the publicly available LivDet 2015 dataset and show that the proposed approach leads to robustness to fingerprint spoofs of the target material.

Fingerprint Spoof Generalization

We present a style-transfer based wrapper, called Universal Material Generator (UMG), to improve the generalization performance of any fingerprint spoof detector against spoofs made from materials not seen during training. Specifically, we transfer the style (texture) characteristics between fingerprint images of known materials with the goal of synthesizing fingerprint images corresponding to unknown materials, that may occupy the space between the known materials in the deep feature space. Synthetic live fingerprint images are also added to the training dataset to force the CNN to learn generative-noise invariant features which discriminate between lives and spoofs. The proposed approach is shown to improve the generalization performance of a state-of-the-art spoof detector, namely Fingerprint Spoof Buster, from TDR of 75.24% to 91.78% @ FDR = 0.2%. These results are based on a large-scale dataset of 5,743 live and 4,912 spoof images fabricated using 12 different materials. Additionally, the UMG wrapper is shown to improve the average cross-sensor spoof detection performance from 67.60% to 80.63% when tested on the LivDet 2017 dataset. Training the UMG wrapper requires only 100 live fingerprint images from the target sensor, alleviating the time and resources required to generate large-scale live and spoof datasets for a new sensor. We also fabricate physical spoof artifacts using a mixture of known spoof materials to explore the role of cross-material style transfer in improving generalization performance.

Finding Missing Children: Aging Deep Face Features

Given a gallery of face images of missing children, state-of-the-art face recognition systems fall short in identifying a child (probe) recovered at a later age. We propose an age-progression module that can age-progress deep face features output by any commodity face matcher. For time lapses larger than 10 years (the missing child is found after 10 or more years), the proposed age-progression module improves the closed-set identification accuracy of FaceNet from 40% to 49.56% and CosFace from 56.88% to 61.25% on a child celebrity dataset, namely ITWCC. The proposed method also outperforms state-of-the-art approaches with a rank-1 identification rate from 94.91% to 95.91% on a public aging dataset, FG-NET, and from 99.50% to 99.58% on CACD-VS. These results suggest that aging face features enhances the ability to identify young children who are possible victims of child trafficking or abduction.

DebFace: De-biasing Face Recognition

We address the problem of bias in automated face recognition algorithms, where errors are consistently lower on certain cohorts belonging to specific demographic groups. We present a novel de-biasing adversarial network that learns to extract disentangled feature representations for both unbiased face recognition and demographics estimation. The proposed network consists of one identity classifier and three demographic classifiers (for gender, age, and race) that are trained to distinguish identity and demographic attributes, respectively. Adversarial learning is adopted to minimize correlation among feature factors so as to abate bias influence from other factors. We also design a new scheme to combine demographics with identity features to strengthen robustness of face representation in different demographic groups. The experimental results show that our approach is able to reduce bias in face recognition as well as demographics estimation while achieving state-of-the-art performance.

Aging Deep Face Features: Finding Missing Children

Given a gallery of face images of missing children, state-of-the-art face recognition systems fall short in identifying a child (probe) recovered at a later age. We propose an age-progression module that can age-progress deep face features output by any commodity face matcher. For time lapses larger than 10 years (the missing child is found after 10 or more years), the proposed age-progression module improves the closed-set identification accuracy of FaceNet from 40% to 49.56% and CosFace from 56.88% to 61.25% on a child celebrity dataset, namely ITWCC. The proposed method also outperforms state-of-the-art approaches with a rank-1 identification rate from 94.91% to 95.91% on a public aging dataset, FG-NET, and from 99.50% to 99.58% on CACD-VS. These results suggest that aging face features enhances the ability to identify young children who are possible victims of child trafficking or abduction.