Uncertainty in the predictions of learning enabled components hinders their deployment safety-critical cyber-physical systems (CPS). A shift from training distribution a component (LEC) is one source uncertainty LEC's predictions. Detection this or out-of-distribution (OOD) detection on individual datapoints has therefore gained attention recently. But many applications, inputs to CPS form temporal sequence. Existing techniques for OOD time-series data either do not exploit relationships...

As machine learning models continue to achieve impressive performance across different tasks, the importance of effective anomaly detection for such has increased as well. It is common knowledge that even well-trained lose their ability function effectively on out-of-distribution inputs. Thus, (OOD) received some attention recently. In vast majority cases, it uses distribution estimated by training dataset OOD detection. We demonstrate current detectors inherit biases in dataset,...

Incorporating learning based components in the current state-of-the-art cyber-physical systems (CPS) has been a challenge due to brittleness of underlying deep neural networks. On bright side, if executed correctly with safety guarantees, this ability revolutionize domains like autonomous systems, medicine, and other safety-critical domains. This is because it would allow system designers use high-dimensional outputs from sensors camera LiDAR. The trepidation deploying vision LiDAR comes...

The use of learning based components in cyber-physical systems (CPS) has created a gamut possible avenues to high dimensional real world signals generated from sensors like camera and LiDAR. ability process such can be largely attributed the adoption high-capacity function approximators deep neural networks. However, this does not come without its potential perils. pitfalls arise over-fitting, subsequent unsafe behavior when exposed unknown environments. One challenge is that, input spaces...

Deep neural network (DNN) models have proven to be vulnerable adversarial digital and physical attacks. In this paper, we propose a novel attack- dataset-agnostic real-time detector for both types of inputs DNN-based perception systems. particular, the proposed relies on observation that images are sensitive certain label-invariant transformations. Specifically, determine if an image has been adversarially manipulated, checks output target classifier given input changes significantly after...

As designs grow in size and complexity, design verification becomes one of the most difficult costly tasks facing teams. Formal techniques offer great promise because their ability to exhaustively explore behaviors. However, formal also have a reputation for being labor-intensive limited small blocks. Is there any hope successful application at scale? We answer this question affirmatively by digging deeper understand what real technological issues opportunities are. First, we look...

Uncertainty in the predictions of learning-enabled components hinders their deployment safety-critical cyber-physical systems (CPS). A shift from training distribution a component (LEC) is one source uncertainty LEC’s predictions. Detection this or out-of-distribution (OOD) detection on individual datapoints has therefore gained attention recently. But many applications, inputs to CPS form temporal sequence. Existing techniques for OOD time-series data either do not exploit relationships...

Recently it has been shown that state-of-the-art NLP models are vulnerable to adversarial attacks, where the predictions of a model can be drastically altered by slight modifications input (such as synonym substitutions). While several defense techniques have proposed, and adapted, discrete nature text benefits general-purpose regularization methods such label smoothing for language models, not studied. In this paper, we study robustness provided strategies in foundational diverse tasks both...

The performance of machine learning models can significantly degrade under distribution shifts the data. We propose a new method for classification which improve robustness to shifts, by combining expert knowledge about ``high-level" structure data with standard classifiers. Specifically, we introduce two-stage classifiers called memory First, these identify prototypical points -- memories cluster training This step is based on features designed guidance; instance, image they be extracted...

Abstract Pulse-coupled neural network (PCNN) model is widely used in digital image processing, but it always a difficult problem to set parameters and determine the optimal segmentation. By analyzing firing characteristics setting for non-coupled linking PCNN, we propose an improved PCNN The introduce coupling effect of neighboring neurons into dynamic threshold subsystem, using combination DNN network, manual adjusting on step length initial value. When value adjusted properly, segmentation...

Cyber-physical systems (CPS) with learning-enabled components suffer from reduced performance under distribution shift. In this paper, we consider the problem of motion prediction within an autonomous racing setting. such a setting, ability to predict adversaries' behavior is essential for safe and efficient planning. We propose method using memories detect anomalous input incrementally learn model online, quickly adapt unseen behaviors. our experiments, demonstrate effectiveness approach in...

Deep neural networks (DNNs) have the ability to transform inference with medical data. However, large DNN models need sufficient labeled data be effective at generalization. This involves considerable manual efforts generate quality in copious amounts for hungry learning tasks. Data programming addresses this issue by using weak labeling functions obtained from experts label unlabeled Such now become a key component pipeline. A step further direction is ask question: it possible...

BabyBERTa, a language model trained on small-scale child-directed speech while none of the words are unmasked during training, has been shown to achieve level grammaticality comparable that RoBERTa-base, which is 6,000 times more and 15 parameters. Relying this promising result, we explore in paper performance BabyBERTa-based models downstream tasks, focusing Semantic Role Labeling (SRL) two Extractive Question Answering with aim building efficient systems rely less data smaller models. We...

Optical fiber laser hydrophone which has more research potentiality and value owing to higher acoustic pressure sensitivity, smaller size lower difficulty of multiplexing. However the detecting capacity low frequency signal optical will be limited because noise such as 1/f thermal pumped laser. In order suppress these noises, iterative discrete wavelet transformation algorithm was designed used multi-scale trait transform. The different spectral components underwater were separated noises...

Deep neural network (DNN) models have proven to be vulnerable adversarial digital and physical attacks. In this paper, we propose a novel attack- dataset-agnostic real-time detector for both types of inputs DNN-based perception systems. particular, the proposed relies on observation that images are sensitive certain label-invariant transformations. Specifically, determine if an image has been adversarially manipulated, checks output target classifier given input changes significantly after...

Nowadays, due to some internal conflicts such as political and economic problems, especially religional intervention by other great powers, wars racial oppression have occured even lead a large number of citizens turing be refugees, finally here comes global refugee problems.In this paper, the author, through method literature review, focues on relationship between international refugees human rights, finds out available ways solve it extent.Through research, author there are several kind...

Recently it has been shown that state-of-the-art NLP models are vulnerable to adversarial attacks, where the predictions of a model can be drastically altered by slight modifications input (such as synonym substitutions). While several defense techniques have proposed, and adapted, discrete nature text benefits general-purpose regularization methods such label smoothing for language models, not studied. In this paper, we study robustness provided various strategies in foundational diverse...