arXiv subjects

W. Ronny Huang

Publications and source records attributed to W. Ronny Huang.

25 records · Page 2Linked to original sources

arXiv2020

DeepErase: Weakly Supervised Ink Artifact Removal in Document Text Images

Paper-intensive industries like insurance, law, and government have long leveraged optical character recognition (OCR) to automatically transcribe hordes of scanned documents into text strings for downstream processing. Even in 2019, there are still many scanned documents and mail that come into businesses in non-digital format. Text to be extracted from real world documents is often nestled inside rich formatting, such as tabular structures or forms with fill-in-the-blank boxes or underlines whose ink often touches or even strikes through the ink of the text itself. Further, the text region could have random ink smudges or spurious strokes. Such ink artifacts can severely interfere with the performance of recognition algorithms or other downstream processing tasks. In this work, we propose DeepErase, a neural-based preprocessor to erase ink artifacts from text images. We devise a method to programmatically assemble real text images and real artifacts into realistic-looking "dirty" text images, and use them to train an artifact segmentation network in a weakly supervised manner, since pixel-level annotations are automatically obtained during the assembly process. In addition to high segmentation accuracy, we show that our cleansed images achieve a significant boost in recognition accuracy by popular OCR software such as Tesseract 4.0. Finally, we test DeepErase on out-of-distribution datasets (NIST SDB) of scanned IRS tax return forms and achieve double-digit improvements in accuracy. All experiments are performed on both printed and handwritten text. Code for all experiments is available at https://github.com/yikeqicn/DeepErase

cs.CV↗

arXiv2019

Transferable Clean-Label Poisoning Attacks on Deep Neural Nets

Clean-label poisoning attacks inject innocuous looking (and "correctly" labeled) poison images into training data, causing a model to misclassify a targeted image after being trained on this data. We consider transferable poisoning attacks that succeed without access to the victim network's outputs, architecture, or (in some cases) training data. To achieve this, we propose a new "polytope attack" in which poison images are designed to surround the targeted image in feature space. We also demonstrate that using Dropout during poison creation helps to enhance transferability of this attack. We achieve transferable attack success rates of over 50% while poisoning only 1% of the training set.

stat.ML↗

arXiv2018

Accurate, Data-Efficient Learning from Noisy, Choice-Based Labels for Inherent Risk Scoring

Inherent risk scoring is an important function in anti-money laundering, used for determining the riskiness of an individual during onboarding $\textit{before}$ fraudulent transactions occur. It is, however, often fraught with two challenges: (1) inconsistent notions of what constitutes as high or low risk by experts and (2) the lack of labeled data. This paper explores a new paradigm of data labeling and data collection to tackle these issues. The data labeling is choice-based; the expert does not provide an absolute risk score but merely chooses the most/least risky example out of a small choice set, which reduces inconsistency because experts make only relative judgments of risk. The data collection is synthetic; examples are crafted using optimal experimental design methods, obviating the need for real data which is often difficult to obtain due to regulatory concerns. We present the methodology of an end-to-end inherent risk scoring algorithm that we built for a large financial institution. The system was trained on a small set of synthetic data (188 examples, 24 features) whose labels are obtained via the choice-based paradigm using an efficient number of expert labelers. The system achieves 89% accuracy on a test set of 52 examples, with an area under the ROC curve of 93%.

cs.LG↗

arXiv2018

Poison Frogs! Targeted Clean-Label Poisoning Attacks on Neural Networks

Data poisoning is an attack on machine learning models wherein the attacker adds examples to the training set to manipulate the behavior of the model at test time. This paper explores poisoning attacks on neural nets. The proposed attacks use "clean-labels"; they don't require the attacker to have any control over the labeling of training data. They are also targeted; they control the behavior of the classifier on a $\textit{specific}$ test instance without degrading overall classifier performance. For example, an attacker could add a seemingly innocuous image (that is properly labeled) to a training set for a face recognition engine, and control the identity of a chosen person at test time. Because the attacker does not need to control the labeling function, poisons could be entered into the training set simply by leaving them on the web and waiting for them to be scraped by a data collection bot. We present an optimization-based method for crafting poisons, and show that just one single poison image can control classifier behavior when transfer learning is used. For full end-to-end training, we present a "watermarking" strategy that makes poisoning reliable using multiple ($\approx$50) poisoned training instances. We demonstrate our method by generating poisoned frog images from the CIFAR dataset and using them to manipulate image classifiers.

cs.LG↗

arXiv2016

Terahertz-driven, all-optical electron gun

Ultrashort electron beams with narrow energy spread, high charge, and low jitter are essential for resolving phase transitions in metals, semiconductors, and molecular crystals. These semirelativistic beams, produced by phototriggered electron guns, are also injected into accelerators for x-ray light sources. The achievable resolution of these time-resolved electron diffraction or x-ray experiments has been hindered by surface field and timing jitter limitations in conventional RF guns, which thus far are <200 MV/m and >96 fs, respectively. A gun driven by optically-generated single-cycle THz pulses provides a practical solution to enable not only GV/m surface fields but also absolute timing stability, since the pulses are generated by the same laser as the phototrigger. Here, we demonstrate an all-optical THz gun yielding peak electron energies approaching 1 keV, accelerated by 300 MV/m THz fields in a novel micron-scale waveguide structure. We also achieve quasimonoenergetic, sub-keV bunches with 32 fC of charge, which can already be used for time-resolved low-energy electron diffraction. Such ultracompact, easy to implement guns driven by intrinsically synchronized THz pulses that are pumped by an amplified arm of the already present photoinjector laser provide a new tool with potential to transform accelerator based science.

physics.acc-ph↗

arXiv2015

THz generation using a reflective stair-step echelon

We present a novel method for THz generation in lithium niobate using a reflective stair-step echelon structure. The echelon produces a discretely tilted pulse front with less angular dispersion compared to a high groove-density grating. The THz output was characterized using both a 1-lens and 3-lens imaging system to set the tilt angle at room and cryogenic temperatures. Using broadband 800 nm pulses with a pulse energy of 0.95 mJ and a pulse duration of 70 fs (24 nm FWHM bandwidth, 39 fs transform limited width), we produced THz pulses with field strengths as high as 500 kV/cm and pulse energies as high as 3.1 $μ$J. The highest conversion efficiency we obtained was 0.33%. In addition, we find that the echelon is easily implemented into an experimental setup for quick alignment and optimization.

physics.optics↗

arXiv2015

Toward a terahertz-driven electron gun

Femtosecond electron bunches with keV energies and eV energy spread are needed by condensed matter physicists to resolve state transitions in carbon nanotubes, molecular structures, organic salts, and charge density wave materials. These semirelativistic electron sources are not only of interest for ultrafast electron diffraction, but also for electron energy-loss spectroscopy and as a seed for x-ray FELs. Thus far, the output energy spread (hence pulse duration) of ultrafast electron guns has been limited by the achievable electric field at the surface of the emitter, which is 10 MV/m for DC guns and 200 MV/m for RF guns. A single-cycle THz electron gun provides a unique opportunity to not only achieve GV/m surface electric fields but also with relatively low THz pulse energies, since a single-cycle transform-limited waveform is the most efficient way to achieve intense electric fields. Here, electron bunches of 50 fC from a flat copper photocathode are accelerated from rest to tens of eV by a microjoule THz pulse with peak electric field of 72 MV/m at 1 kHz repetition rate. We show that scaling to the readily-available GV/m THz field regime would translate to monoenergetic electron beams of ~100 keV.

physics.optics↗