Injecting around 50 poisoned samples with a stealthy trigger creates backdoors in deep learning models achieving over 90% attack success under a weak threat model with no model or data knowledge required.
hub
Deep residual learning for image recognition
17 Pith papers cite this work. Polarity classification is still indexing.
17
Pith papers citing it
hub tools
citation-role summary
background 2
citation-polarity summary
roles
background 2polarities
background 2representative citing papers
Learned iterative methods based on gradient descent, Gauss-Newton, and Quasi-Newton updates are applied to quantitative photoacoustic tomography, showing improved generalization on simulated and digital twin data with scarce training data and modeling errors.
Synthetic data augmentation improves instance segmentation performance for chicken carcasses when real annotated data is limited.
FreshPRINCE and DrCIF, two new unsupervised feature-based regressors adapted from time series classification, significantly outperform other methods on an expanded archive of 63 TSER problems and are the only ones to beat rotation forest by a statistically significant margin.
The paper shows that multiple-identity image attacks succeed due to modest angular separation between matching (~90°) and non-matching (40-60°) face representations, with image morphing and representation inversion realizing effective attacks that transfer across comparators.
Kernel interpolation with a constant multiplier scales convolution and fully-connected layers in neural networks to higher resolutions or dimensions without training, producing competitive results on Stable Diffusion and other models.
A LoRA-based residual feature alignment method for efficient machine unlearning on pre-trained models by targeting zero residuals on retained data and shifted residuals on unlearned data.
PSI is a benchmark dataset for pedestrian intention prediction, driver decision modeling, and reasoning generation in traffic interactions, enriched with human textual explanations.
A context-aware CNN using 1792x1792 images and spatial feature aggregation outperforms patch-based methods for colorectal cancer grading by 3.61%.
AirwayNet and BifurcationNet, trained only on CT-derived simulations, localize bronchoscope pose in phantom and cadaver videos with AUC-PR up to 0.997 and drive a robot to four targets at 95% success.
Two data selection techniques (GMM visual similarity and bounding-box diversity) reduce required weakly labeled images by up to 100x on Open Images and 20x on Cityscapes while maintaining semantic segmentation performance.
An experiment with 276 participants finds that vision language model assistance improves human game testers' defect identification, especially with design documentation, while AI errors create challenges.
Flemme is a modular platform separating encoders (conv/transformer/SSM) from encoder-decoder architectures for medical images, with a hierarchical pyramid loss yielding reported average gains of 5.6% Dice and 5.57% PSNR.
Modifying capsule networks to use dynamic routing for intermediate equivariant features instead of output class capsules yields faster training and higher accuracy on multi-class problems.
A DenseNet201 base model trained on a constructed plant leaf disease dataset outperforms baselines and enables faster, more robust transfer learning with less data than general models.
A comparison of FCNN architectures for monocular depth estimation yields a model suitable for real-time operation on NVidia Jetson hardware with evaluation in vSLAM.
citing papers explorer
-
Targeted Backdoor Attacks on Deep Learning Systems Using Data Poisoning
Injecting around 50 poisoned samples with a stealthy trigger creates backdoors in deep learning models achieving over 90% attack success under a weak threat model with no model or data knowledge required.