Protecting against simultaneous data poisoning attacks

Current backdoor defense methods are evaluated against a single attack at a time. This is unrealistic, as powerful machine learning systems are trained on large datasets scraped from the internet, which may be attacked multiple times by one or more attackers. We demonstrate that simultaneously executed data poisoning attacks can effectively install multiple backdoors in a single model without substantially degrading clean accuracy. Furthermore, we show that existing backdoor defense methods do not effectively prevent attacks in this setting. Finally, we leverage insights into the nature of backdoor attacks to develop a new defense, BaDLoss, that is effective in the multi-attack setting. With minimal clean accuracy degradation, BaDLoss attains an average attack success rate in the multi-attack setting of 7.98% in CIFAR-10 and 10.29% in GTSRB, compared to the average of other defenses at 64.48% and 84.28% respectively.

A deeper look at depth pruning of LLMs

Large Language Models (LLMs) are not only resource-intensive to train but even more costly to deploy in production. Therefore, recent work has attempted to prune blocks of LLMs based on cheap proxies for estimating block importance, effectively removing 10% of blocks in well-trained LLaMa-2 and Mistral 7b models without any significant degradation of downstream metrics. In this paper, we explore different block importance metrics by considering adaptive metrics such as Shapley value in addition to static ones explored in prior work. We show that adaptive metrics exhibit a trade-off in performance between tasks i.e., improvement on one task may degrade performance on the other due to differences in the computed block influences. Furthermore, we extend this analysis from a complete block to individual self-attention and feed-forward layers, highlighting the propensity of the self-attention layers to be more amendable to pruning, even allowing removal of upto 33% of the self-attention layers without incurring any performance degradation on MMLU for Mistral 7b (significant reduction in costly maintenance of KV-cache). Finally, we look at simple performance recovery techniques to emulate the pruned layers by training lightweight additive bias or low-rank linear adapters. Performance recovery using emulated updates avoids performance degradation for the initial blocks (up to 5% absolute improvement on MMLU), which is either competitive or superior to the learning-based technique.

Investigating the Nature of 3D Generalization in Deep Neural Networks

Visual object recognition systems need to generalize from a set of 2D training views to novel views. The question of how the human visual system can generalize to novel views has been studied and modeled in psychology, computer vision, and neuroscience. Modern deep learning architectures for object recognition generalize well to novel views, but the mechanisms are not well understood. In this paper, we characterize the ability of common deep learning architectures to generalize to novel views. We formulate this as a supervised classification task where labels correspond to unique 3D objects and examples correspond to 2D views of the objects at different 3D orientations. We consider three common models of generalization to novel views: (i) full 3D generalization, (ii) pure 2D matching, and (iii) matching based on a linear combination of views. We find that deep models generalize well to novel views, but they do so in a way that differs from all these existing models. Extrapolation to views beyond the range covered by views in the training set is limited, and extrapolation to novel rotation axes is even more limited, implying that the networks do not infer full 3D structure, nor use linear interpolation. Yet, generalization is far superior to pure 2D matching. These findings help with designing datasets with 2D views required to achieve 3D generalization. Code to reproduce our experiments is publicly available: https://github.com/shoaibahmed/investigating_3d_generalization.git

Blockwise Self-Supervised Learning at Scale

Current state-of-the-art deep networks are all powered by backpropagation. In this paper, we explore alternatives to full backpropagation in the form of blockwise learning rules, leveraging the latest developments in self-supervised learning. We show that a blockwise pretraining procedure consisting of training independently the 4 main blocks of layers of a ResNet-50 with Barlow Twins' loss function at each block performs almost as well as end-to-end backpropagation on ImageNet: a linear probe trained on top of our blockwise pretrained model obtains a top-1 classification accuracy of 70.48%, only 1.1% below the accuracy of an end-to-end pretrained network (71.57% accuracy). We perform extensive experiments to understand the impact of different components within our method and explore a variety of adaptations of self-supervised learning to the blockwise paradigm, building an exhaustive understanding of the critical avenues for scaling local learning rules to large networks, with implications ranging from hardware design to neuroscience.

Domain Generalization for Robust Model-Based Offline Reinforcement Learning

Existing offline reinforcement learning (RL) algorithms typically assume that training data is either: 1) generated by a known policy, or 2) of entirely unknown origin. We consider multi-demonstrator offline RL, a middle ground where we know which demonstrators generated each dataset, but make no assumptions about the underlying policies of the demonstrators. This is the most natural setting when collecting data from multiple human operators, yet remains unexplored. Since different demonstrators induce different data distributions, we show that this can be naturally framed as a domain generalization problem, with each demonstrator corresponding to a different domain. Specifically, we propose Domain-Invariant Model-based Offline RL (DIMORL), where we apply Risk Extrapolation (REx) (Krueger et al., 2020) to the process of learning dynamics and rewards models. Our results show that models trained with REx exhibit improved domain generalization performance when compared with the natural baseline of pooling all demonstrators' data. We observe that the resulting models frequently enable the learning of superior policies in the offline model-based RL setting, can improve the stability of the policy learning process, and potentially enable increased exploration.

Metadata Archaeology: Unearthing Data Subsets by Leveraging Training Dynamics

Modern machine learning research relies on relatively few carefully curated datasets. Even in these datasets, and typically in `untidy' or raw data, practitioners are faced with significant issues of data quality and diversity which can be prohibitively labor intensive to address. Existing methods for dealing with these challenges tend to make strong assumptions about the particular issues at play, and often require a priori knowledge or metadata such as domain labels. Our work is orthogonal to these methods: we instead focus on providing a unified and efficient framework for Metadata Archaeology -- uncovering and inferring metadata of examples in a dataset. We curate different subsets of data that might exist in a dataset (e.g. mislabeled, atypical, or out-of-distribution examples) using simple transformations, and leverage differences in learning dynamics between these probe suites to infer metadata of interest. Our method is on par with far more sophisticated mitigation methods across different tasks: identifying and correcting mislabeled examples, classifying minority-group samples, prioritizing points relevant for training and enabling scalable human auditing of relevant examples.

Revisiting the Shape-Bias of Deep Learning for Dermoscopic Skin Lesion Classification

It is generally believed that the human visual system is biased towards the recognition of shapes rather than textures. This assumption has led to a growing body of work aiming to align deep models' decision-making processes with the fundamental properties of human vision. The reliance on shape features is primarily expected to improve the robustness of these models under covariate shift. In this paper, we revisit the significance of shape-biases for the classification of skin lesion images. Our analysis shows that different skin lesion datasets exhibit varying biases towards individual image features. Interestingly, despite deep feature extractors being inclined towards learning entangled features for skin lesion classification, individual features can still be decoded from this entangled representation. This indicates that these features are still represented in the learnt embedding spaces of the models, but not used for classification. In addition, the spectral analysis of different datasets shows that in contrast to common visual recognition, dermoscopic skin lesion classification, by nature, is reliant on complex feature combinations beyond shape-bias. As a natural consequence, shifting away from the prevalent desire of shape-biasing models can even improve skin lesion classifiers in some cases.

Improving Health Mentioning Classification of Tweets using Contrastive Adversarial Training

Health mentioning classification (HMC) classifies an input text as health mention or not. Figurative and non-health mention of disease words makes the classification task challenging. Learning the context of the input text is the key to this problem. The idea is to learn word representation by its surrounding words and utilize emojis in the text to help improve the classification results. In this paper, we improve the word representation of the input text using adversarial training that acts as a regularizer during fine-tuning of the model. We generate adversarial examples by perturbing the embeddings of the model and then train the model on a pair of clean and adversarial examples. Additionally, we utilize contrastive loss that pushes a pair of clean and perturbed examples close to each other and other examples away in the representation space. We train and evaluate the method on an extended version of the publicly available PHM2017 dataset. Experiments show an improvement of 1.0% over BERT-Large baseline and 0.6% over RoBERTa-Large baseline, whereas 5.8% over the state-of-the-art in terms of F1 score. Furthermore, we provide a brief analysis of the results by utilizing the power of explainable AI.

Identifying Layers Susceptible to Adversarial Attacks

Common neural network architectures are susceptible to attack by adversarial samples. Neural network architectures are commonly thought of as divided into low-level feature extraction layers and high-level classification layers; susceptibility of networks to adversarial samples is often thought of as a problem related to classification rather than feature extraction. We test this idea by selectively retraining different portions of VGG and ResNet architectures on CIFAR-10, Imagenette and ImageNet using non-adversarial and adversarial data. Our experimental results show that susceptibility to adversarial samples is associated with low-level feature extraction layers. Therefore, retraining high-level layers is insufficient for achieving robustness. This phenomenon could have two explanations: either, adversarial attacks yield outputs from early layers that are indistinguishable from features found in the attack classes, or adversarial attacks yield outputs from early layers that differ statistically from features for non-adversarial samples and do not permit consistent classification by subsequent layers. We test this question by large-scale non-linear dimensionality reduction and density modeling on distributions of feature vectors in hidden layers and find that the feature distributions between non-adversarial and adversarial samples differ substantially. Our results provide new insights into the statistical origins of adversarial samples and possible defenses.

Reliable Model Compression via Label-Preservation-Aware Loss Functions

Model compression is a ubiquitous tool that brings the power of modern deep learning to edge devices with power and latency constraints. The goal of model compression is to take a large reference neural network and output a smaller and less expensive compressed network that is functionally equivalent to the reference. Compression typically involves pruning and/or quantization, followed by re-training to maintain the reference accuracy. However, it has been observed that compression can lead to a considerable mismatch in the labels produced by the reference and the compressed models, resulting in bias and unreliability. To combat this, we present a framework that uses a teacher-student learning paradigm to better preserve labels. We investigate the role of additional terms to the loss function and show how to automatically tune the associated parameters. We demonstrate the effectiveness of our approach both quantitatively and qualitatively on multiple compression schemes and accuracy recovery algorithms using a set of 8 different real-world network architectures. We obtain a significant reduction of up to 4.1X in the number of mismatches between the compressed and reference models, and up to 5.7X in cases where the reference model makes the correct prediction.