Safe deep neural networks for autonomous driving vehicles

SafeDNN for safe deep neural networks

Integrated framework for the dependability evaluation of deep neural networks in autonomous cars

We have published 10+ peer- reviewed publications in the area of safe and reliable ML at top-tier venues journals in computer vision, machine learning, and safety, including CVPR, ECAI and SAFECOMP

We develop a SafeDNN framework, which can automatically identify failures in your machine learning model, especially in deep neural networks and test robustness of your model against adversarial attacks

We also developed a tool in compliance with ASPICE standard to automatically detect noisy labels in image recognition and object detection datasets. Our development process meets the functional safety requirements for the confidence in the use of software tools according to ISO 26262-8:2018, clause 11 section for an offline support tool with a tool confidence level up to TCL3

We know that safety of deep neural networks assessment is not sufficiently covered by existing functional safety standards such as ISO 26262. IEC 61508 and ISO/PAS 21448 (SOTIF) standards, but our team build internal knowledge how to ensure safe operation. We also offer SOTIF trainings with special treatment of deep neural network safety

Learn more about our projects and let's organize a call how our team can help you to ensure safety of deep neural networks. We can help you to ensure safety and security through AI assessment services

SOTIF training

SafeDNN project has begun

Today, we have started SafeDNN project funded by The National Centre for Research and Development. Our goal is to develop software that will support the development and deployment of deep neural networks in safety-critical systems, especially in autonomous vehicles. The project will last for the next 36 months.

SafeDNN for object detection problems

This year we will focus on extension of the currently developed methods to adjust them for state-of-the-art object detectors that use deep neural networks. Out target is to develop the largest framework that will support safe deployment of deep neural networks in safety-critical systems, especially autonomous driving applications.

Neuron activation patterns on CVPR

Our paper “Detection of out-of-distribution samples using binary neuron activation patterns” was presented on CVPR 2023 conference. More information at: https://safednn.com/nap/

ISO/PAS 21448:2022 – SOTIF standard

ISO/PAS 21448:2022 also known as SOTIF (Safety of the Intended Functionality) is a safety standard that addresses hazards and risks associated with the functioning of automotive systems when they are operating as intended, but may still lead to accidents or injuries due to unforeseen systematic failures or external factors. SOTIF focuses on the potential limitations […]

ALL NEWS

Team

Defending against sparse adversarial attacks using impulsive noise reduction filters

K. Radlak, M. Szczepankiewicz, B. Smolka

Abstract Deep Neural Networks (DNNs) have been deployed in many real-world applications in various domains, both industry and academic, and have proven to deliver outstanding performance. However, DNNs are vulnerable to adversarial attacks, that are small perturbations embedded in an image. As a result, the introduction of DNNs into safety-critical systems, such as autonomous vehicles, […]

Improving the detection of noisy labels in image datasets using MultiNET Flexible Confidence Learning

A. Popowicz; K. Radlak; S. Lasota; K. Szczepankiewicz; M. Szczepankiewicz

The effectiveness of machine learning algorithms depends on proper preparation of the training dataset. Erroneously labeled images in the training data will degrade algorithmic efficiency and cause unpredictable model behavior. Verifying labels in the numerous available databases remains a complicated and laborious task. In this article, we present a MultiNET approach that allows (...)

Testing and Verification of the Deep Neural Networks Against Sparse Pixel Defects

Michal Szczepankiewicz, Krystian Radlak, Karolina Szczepankiewicz, Adam Popowicz & Pawel Zawistowski

Deep neural networks can produce outstanding results when applied to image recognition tasks but are susceptible to image defects and modifications. This work introduces a novel sparse adversarial attack generation method based on differential evolution strategy.

Combating Label Noise in Image Data Using MultiNET Flexible Confident Learning

A. Popowicz, K. Radlak, S. Lasota, K. Szczepankiewicz, M. Szczepankiewicz

Erroneously labeled training images can degrade the final accuracy and additionally lead to unpredictable model behavior, reducing reliability. In this paper, we propose MultiNET, a novel method for the automatic detection of noisy labels within image datasets. MultiNET is an adaptation of the current state-of-the-art confident learning method (...)

All publications

SOTIF training

SafeDNN project has begun

SafeDNN for object detection problems

Neuron activation patterns on CVPR

ISO/PAS 21448:2022 – SOTIF standard

News

Karolina Szczepankiewicz

Prof. Adam Popowicz

Michał Szczepankiewicz

Dr. Krystian Radlak

Bartłomiej Olber

Krystian Chachuła

Piotr Satała

Team

Defending against sparse adversarial attacks using impulsive noise reduction filters

Improving the detection of noisy labels in image datasets using MultiNET Flexible Confidence Learning

Testing and Verification of the Deep Neural Networks Against Sparse Pixel Defects

Combating Label Noise in Image Data Using MultiNET Flexible Confident Learning

Publications

Contact details: