TREND: Transferability-Based Robust ENsemble Design

نویسندگان

چکیده

Deep learning models hold state-of-the-art performance in many fields, but their vulnerability to adversarial examples poses a threat ubiquitous deployment practical settings. Additionally, inputs generated on one classifier have been shown transfer other classifiers trained similar data, which makes the attacks possible even if model parameters are not revealed adversary. This property of transferability has yet systematically studied, leading gap our understanding robustness neural networks inputs. In this work, we study effect network architecture, optimizer, input, weight, and activation quantization samples. We also different attacks. Our experiments reveal that is significantly hampered by input architectural mismatch between source target, choice optimizer turns out be critical. observe architecture-dependent for both weight quantized models. To quantify transferability, use simple metric demonstrate utility designing methodology build ensembles with improved robustness. When attacking “gradient domination” single ensemble member hampers existing combat propose new attack. compare proposed attack techniques show its effectiveness. Finally, an consisting carefully chosen diverse achieves better than would otherwise network. The code work made available at https://github.com/purdue-nrl/TREND .

برای دانلود باید عضویت طلایی داشته باشید

برای دانلود متن کامل این مقاله و بیش از 32 میلیون مقاله دیگر ابتدا ثبت نام کنید

اگر عضو سایت هستید لطفا وارد حساب کاربری خود شوید

منابع مشابه

Robust ensemble-based multi-objective optimization

• A submitted manuscript is the author's version of the article upon submission and before peer-review. There can be important differences between the submitted version and the official published version of record. People interested in the research are advised to contact the author for the final version of the publication, or visit the DOI to the publisher's website. • The final author version ...

متن کامل

Robust Design and Reliability-based Design Optimization

A large number of problems in manufacturing processes, production planning, finance and engineering design require an understanding of potential sources of variations and quantification of the effect of variations on product behavior and performance. Traditionally, in engineering problems uncertainties have been formulated only through coarse safety factors. Such methods often lead to overdesig...

متن کامل

Robust and Reliability-Based Design

The idea behind this special supplement came out of discussions with the Editor, Dr. Michael McCarthy, about the need to present some recent advances in the area of risk-based and robust design. We felt, and still feel, that there is an ever increasing need to design mechanical and structural systems that are risk tolerant and robust. It was our belief that it would be useful if, by way of a de...

متن کامل

Classifier Ensemble Framework: a Diversity Based Approach

Pattern recognition systems are widely used in a host of different fields. Due to some reasons such as lack of knowledge about a method based on which the best classifier is detected for any arbitrary problem, and thanks to significant improvement in accuracy, researchers turn to ensemble methods in almost every task of pattern recognition. Classification as a major task in pattern recognition,...

متن کامل

ذخیره در منابع من


  با ذخیره ی این منبع در منابع من، دسترسی به آن را برای استفاده های بعدی آسان تر کنید

ژورنال

عنوان ژورنال: IEEE transactions on artificial intelligence

سال: 2023

ISSN: ['2691-4581']

DOI: https://doi.org/10.1109/tai.2022.3175172