Backdoor attacks have been shown to be a serious threat against deep learning systems such as biometric authentication and autonomous driving. An effective backdoor attack could enforce the model misbehave under certain predefined conditions, i.e., triggers, but behave normally otherwise. However, the triggers of existing attacks are directly injected in the pixel space, which tend to be detectable by existing defenses and visually identifiable at both training and inference stages. In this paper, we propose a new backdoor attack FTROJAN through trojaning the frequency domain. The key intuition is that triggering perturbations in the frequency domain correspond to small pixel-wise perturbations dispersed across the entire image, breaking the underlying assumptions of existing defenses and making the poisoning images visually indistinguishable from clean ones. We evaluate FTROJAN in several datasets and tasks showing that it achieves a high attack success rate without significantly degrading the prediction accuracy on benign inputs. Moreover, the poisoning images are nearly invisible and retain high perceptual quality. We also evaluate FTROJAN against state-of-the-art defenses as well as several adaptive defenses that are designed on the frequency domain. The results show that FTROJAN can robustly elude or significantly degenerate the performance of these defenses.
Pre-Trained Vision-Language Models (VL-PTMs) have shown promising capabilities in grounding natural language in image data, facilitating a broad variety of cross-modal tasks. However, we note that there exists a significant gap between the objective forms of model pre-training and fine-tuning, resulting in a need for large amounts of labeled data to stimulate the visual grounding capability of VL-PTMs for downstream tasks. To address the challenge, we present Cross-modal Prompt Tuning (CPT, alternatively, Colorful Prompt Tuning), a novel paradigm for tuning VL-PTMs, which reformulates visual grounding into a fill-in-the-blank problem with color-based co-referential markers in image and text, maximally mitigating the gap. In this way, CPT enables strong few-shot and even zero-shot visual grounding capabilities of VL-PTMs. Comprehensive experimental results show that the prompt-tuned VL-PTMs outperform their fine-tuned counterparts by a large margin (e.g., 17.3% absolute accuracy improvement, and 73.8% relative standard deviation reduction on average with one shot in RefCOCO evaluation). All the data and codes will be available to facilitate future research.
Federated learning (FL) aims to protect data privacy by cooperatively learning a model without sharing private data among users. For Federated Learning of Deep Neural Network with billions of model parameters, existing privacy-preserving solutions are unsatisfactory. Homomorphic encryption (HE) based methods provide secure privacy protections but suffer from extremely high computational and communication overheads rendering it almost useless in practice . Deep learning with Differential Privacy (DP) was implemented as a practical learning algorithm at a manageable cost in complexity. However, DP is vulnerable to aggressive Bayesian restoration attacks as disclosed in the literature and demonstrated in experimental results of this work. To address the aforementioned perplexity, we propose a novel Bayesian Privacy (BP) framework which enables Bayesian restoration attacks to be formulated as the probability of reconstructing private data from observed public information. Specifically, the proposed BP framework accurately quantifies privacy loss by Kullback-Leibler (KL) Divergence between the prior distribution about the privacy data and the posterior distribution of restoration private data conditioning on exposed information}. To our best knowledge, this Bayesian Privacy analysis is the first to provides theoretical justification of secure privacy-preserving capabilities against Bayesian restoration attacks. As a concrete use case, we demonstrate that a novel federated deep learning method using private passport layers is able to simultaneously achieve high model performance, privacy-preserving capability and low computational complexity. Theoretical analysis is in accordance with empirical measurements of information leakage extensively experimented with a variety of DNN networks on image classification MNIST, CIFAR10, and CIFAR100 datasets.
Recurrent neural networks for language models like long short-term memory (LSTM) have been utilized as a tool for modeling and predicting long term dynamics of complex stochastic molecular systems. Recently successful examples on learning slow dynamics by LSTM are given with simulation data of low dimensional reaction coordinate. However, in this report we show that the following three key factors significantly affect the performance of language model learning, namely dimensionality of reaction coordinates, temporal resolution and state partition. When applying recurrent neural networks to molecular dynamics simulation trajectories of high dimensionality, we find that rare events corresponding to the slow dynamics might be obscured by other faster dynamics of the system, and cannot be efficiently learned. Under such conditions, we find that coarse graining the conformational space into metastable states and removing recrossing events when estimating transition probabilities between states could greatly help improve the accuracy of slow dynamics learning in molecular dynamics. Moreover, we also explore other models like Transformer, which do not show superior performance than LSTM in overcoming these issues. Therefore, to learn rare events of slow molecular dynamics by LSTM and Transformer, it is critical to choose proper temporal resolution (i.e., saving intervals of MD simulation trajectories) and state partition in high resolution data, since deep neural network models might not automatically disentangle slow dynamics from fast dynamics when both are present in data influencing each other.
In recent years, the size of pre-trained language models (PLMs) has grown by leaps and bounds. However, efficiency issues of these large-scale PLMs limit their utilization in real-world scenarios. We present a suite of cost-effective techniques for the use of PLMs to deal with the efficiency issues of pre-training, fine-tuning, and inference. (1) We introduce knowledge inheritance to accelerate the pre-training process by exploiting existing PLMs instead of training models from scratch. (2) We explore the best practice of prompt tuning with large-scale PLMs. Compared with conventional fine-tuning, prompt tuning significantly reduces the number of task-specific parameters. (3) We implement a new inference toolkit, namely InfMoE, for using large-scale PLMs with limited computational resources. Based on our cost-effective pipeline, we pre-train two models: an encoder-decoder bilingual model with 11 billion parameters (CPM-2) and its corresponding MoE version with 198 billion parameters. In our experiments, we compare CPM-2 with mT5 on downstream tasks. Experimental results show that CPM-2 has excellent general language intelligence. Moreover, we validate the efficiency of InfMoE when conducting inference of large-scale models having tens of billions of parameters on a single GPU. All source code and model parameters are available at https://github.com/TsinghuaAI/CPM.
Recent studies show that neural natural language processing (NLP) models are vulnerable to backdoor attacks. Injected with backdoors, models perform normally on benign examples but produce attacker-specified predictions when the backdoor is activated, presenting serious security threats to real-world applications. Since existing textual backdoor attacks pay little attention to the invisibility of backdoors, they can be easily detected and blocked. In this work, we present invisible backdoors that are activated by a learnable combination of word substitution. We show that NLP models can be injected with backdoors that lead to a nearly 100% attack success rate, whereas being highly invisible to existing defense strategies and even human inspections. The results raise a serious alarm to the security of NLP models, which requires further research to be resolved. All the data and code of this paper are released at https://github.com/thunlp/BkdAtk-LWS.