Yunrui Yu

Dr. Yunrui Yu


Research Focus & Metrics

Academic Career Length

5Years (2022-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
13 October 2023 in IEEE Transactions on Pattern Analysis and Machine Intelligence
Verified
Journal Article
Open access
YYYunrui Yu
Yunrui Yu
XGXitong Gao
Xitong Gao
CXCheng-Zhong Xu
Cheng-Zhong Xu
Deep convolutional neural networks (CNNs) can be easily tricked to give incorrect outputs by adding tiny perturbations to the input that are imperceptible to humans. This makes them susceptible to adversarial attacks, and poses significant security risks to deep learning systems, and presents a great challenge in making CNNs robust against such attacks. An influx of defense strategies have thus been proposed to improve the robustness of CNNs. Current attack methods, however, may fail to accurately or efficiently evaluate the robustness of defending models. In this paper, we thus propose a unified $\ell _{p}$ white-box attack strategy, LAFIT, to harness the defender's latent features in its gradient descent steps, and further employ a new loss function to normalize logits to overcome floating-point-based gradient masking. We show that not only is it more efficient, but it is also a stronger adversary than the current state-of-the-art when examined across a wide range of defense mechanisms. This suggests that adversarial attacks/defenses could be contingent on the effective use of the defender's hidden components, and robustness evaluation should no longer view models holistically.

26 July 2024 in Preprints.org
Verified
Preprint
Open access
PDF file
YYYunrui Yu
Yunrui Yu
Newton's kinetic energy equation delineates the laws of motion for objects travelling at speeds much less than the speed of light, an equation extensively validated at these lower velocities. Any new, more complete formulation of kinetic energy must approximate Newton's kinetic energy equation under these conditions. Einstein's special relativity, utilizing the Lorentz factor, introduces a new perspective on spacetime, adhering to the requirement that the Newtonian kinetic energy equation serves as an approximation at low velocities. In this paper, we identify all factor expressions satisfying the condition that their approximate solutions at low velocities align with Newton's kinetic energy equation. We introduce the Yuyunrui factor, incorporating two real number hyperparameters, $\alpha$ and $\beta$, encompassing all potentially correct factors, with the Lorentz factor as a specific instance where $(\alpha=0$, $\beta=1)$. We classify all potentially correct $(\alpha, \beta)$ combinations into six categories based on their distinct scaling effects and present the corresponding scaling effects. The combinations of $(\alpha, \beta)$ values applicable in the real world are necessarily included within these candidates, with the precise values requiring rigorous experimental validation. This work aims to refine our understanding of special relativity beyond the Lorentz factor.

15 November 2022 in arXiv
Verified
Preprint
Open access
PDF file
YYYunrui Yu
Yunrui Yu
XGXitong Gao
Xitong Gao
CXCheng-Zhong Xu
Cheng-Zhong Xu
Adversarial attacks can deceive neural networks by adding tiny perturbations to their input data. Ensemble defenses, which are trained to minimize attack transferability among sub-models, offer a promising research direction to improve robustness against such attacks while maintaining a high accuracy on natural inputs. We discover, however, that recent state-of-the-art (SOTA) adversarial attack strategies cannot reliably evaluate ensemble defenses, sizeably overestimating their robustness. This paper identifies the two factors that contribute to this behavior. First, these defenses form ensembles that are notably difficult for existing gradient-based method to attack, due to gradient obfuscation. Second, ensemble defenses diversify sub-model gradients, presenting a challenge to defeat all sub-models simultaneously, simply summing their contributions may counteract the overall attack objective; yet, we observe that ensemble may still be fooled despite most sub-models being correct. We therefore introduce MORA, a model-reweighing attack to steer adversarial example synthesis by reweighing the importance of sub-model gradients. MORA finds that recent ensemble defenses all exhibit varying degrees of overestimated robustness. Comparing it against recent SOTA white-box attacks, it can converge orders of magnitude faster while achieving higher attack success rates across all ensemble models examined with three different ensemble modes (i.e., ensembling by either softmax, voting or logits). In particular, most ensemble defenses exhibit near or exactly 0% robustness against MORA with $\ell^\infty$ perturbation within 0.02 on CIFAR-10, and 0.01 on CIFAR-100. We make MORA open source with reproducible results and pre-trained models; and provide a leaderboard of ensemble defenses under various attack strategies.

30 September 2024 in Preprints.org
Verified
Preprint
Open access
PDF file
YYYunrui Yu
Yunrui Yu
Newton's classical dynamics accurately describes the motion of objects moving at speeds much less than the speed of light and has been extensively validated under these conditions. Any new dynamics formula must reduce to Newton's equations at low velocities. Einstein's theory of special relativity, which utilizes the Lorentz factor, provides a revised view of space-time and also conforms to Newton's equations for low speeds. However, the Lorentz factor assumes that the maximum anisotropy of the speed of light is zero, a condition not fully supported by experimental evidence. The Michelson-Morley Experiment, conducted at a speed far less than that of light, suggests that while the measured anisotropy is very small, it is not zero. This small value could imply the existence of undiscovered physical laws. To address this, we propose the Yuyunrui factor, which includes two hyperparameters, \(\alpha\) and \(\beta\), encompassing various possible models. In this framework, the Lorentz factor represents a special case with \((\alpha=0, \beta=1)\) and zero anisotropy. We categorize the \((\alpha, \beta)\) combinations into four broad types based on their scaling effects and describe these in detail. Future experiments at higher speeds will be essential to accurately determine the true value of the maximum anisotropy of the speed of light and to refine the proposed factor.

27 September 2024 in Preprints.org
Verified
Preprint
Open access
PDF file
YYYunrui Yu
Yunrui Yu
This paper focuses on the long-standing issue of galactic rotation, i.e. the discrepancy between observed rotation rates and predictions based on Newtonian gravity. While Newton's law holds within the solar system, it fails to explain the higher-than-expected rotation speeds of galaxies, particularly at greater distances from their centers. This study finds that Newton's gravitational formula applies only when gravitational lensing has a negligible impact. We examine the gravitational force on objects farther from the galactic center, where the lensing effect of a supermassive black hole becomes significant. This lensing causes the actual gravitational force to exceed Newtonian predictions, with the discrepancy increasing with distance. This effect offers an alternative explanation for the galactic rotation curve without the need for dark matter. Our findings suggest that Newton's formula requires revision to account for gravitational lensing, a key factor in understanding galactic dynamics and its broader cosmic implications.