Well-trained Model Research Articles

Examination Process Modeling for Intelligent Patent Management: A Multi-aspect Neural Sequential Approach

Recently, the booming growth of patent applications has brought an unprecedented challenge in performing efficient intellectual property management. Therefore, intelligent approaches are urgently needed to analyze intrinsic patterns of patents. However, a long-standing obstacle is the lack of effective methods for modeling the dynamic and diverse examination process of patent applications, which can benefit a wide range of downstream tasks for patent management. In fact, the major challenges lie in how to discover and integrate domain-specific properties from large-scale unlabeled examination data. To this end, in this paper, we propose a Self-supervised Examination Process Modeling (SEPM) framework to learn the contextualized embedding for patents through modeling their examination processes. Specifically, we first design a multi-aspect event embedding layer, which leverages the fine-tuned language model, frequent-pattern embedding, and time encoding to capture the semantic, frequent-pattern and temporal information of examination events, respectively. Then, a mutual-information-aware integration layer is applied to fuse the extracted features into multi-aspect embedding considering their mutual interactions. Further, we develop a multi-objective sequential neural network for learning the contextualized patent representation, which is achieved through jointly learning two self-supervised objectives, namely event code and event lag auto-regression. To explore the application potential of SEPM, we fine-tune the well-trained model for three important downstream tasks of patent management, including the prediction of next events, patent classification, and grant prediction. In the end, extensive experiments with real-world data from the US Patent and Trademark Office verify the effectiveness and application prospects of the proposed framework.

Long short-term memory (LSTM) neural networks for in situ particle velocity determination in material strength experiments under ramp wave compression

In the experiments of measuring the strength of materials under ramp compression, accurately determining in situ particle velocity is crucial for calculating material sound speed during loading–unloading path and materials strength under high pressure. This paper proposes a machine learning approach that utilizes Long Short-Term Memory (LSTM) neural networks and Bayesian optimization algorithms to enhance the analysis of data from ramp compression strength measurement experiments. This method leverages LSTM neural networks to uncover the complex relationship between the rear interface velocity of the sample and the in situ particle velocity in numerical simulations. By using a well-trained network model, it enables direct interpretation of experimental data, leading to accurate predictions of key physical quantities along the loading and unloading paths in ramp compression experiments. A comparative analysis between theoretical curves from numerical simulations and LSTM neural network predictions shows a high degree of consistency. This approach is applied to ramp compression experiments on Ta and CuCrZr materials, demonstrating superior accuracy over the free-surface approximation and incremental impedance matching methods. Additionally, this method relies solely on the equation of state during numerical computations, eliminating the need for the complex constitutive equations required by the transfer function method, thus enhancing data processing efficiency and practicality.

Hierarchical Encoding for Image Inpainting with StyleGAN Inversion

Image inpainting, the process of removing unwanted pixels and seamlessly replacing them with new ones, poses significant challenges requiring algorithms to understand image context and generate realistic replacements. With applications ranging from content generation to image editing, image inpainting has garnered significant interest. Traditional approaches involve training deep neural network models from scratch using binary masks to identify regions for inpainting. Recent advancements have shown the feasibility of leveraging well-trained image generation models, such as StyleGANs, for inpainting tasks. However, effectively embedding images into StyleGAN's latent space and addressing the challenges of diverse inpainting remain key obstacles. In this work, we propose a hierarchical encoder tailored to encode visible and missing features seamlessly. Additionally, we introduce a single-stage architecture capable of encoding both low-rate and high-rate latent features used by StyleGAN. While low-rate latent features offer a comprehensive understanding of images, high-rate latent features excel in transmitting intricate details to the generator. Through extensive experiments, we demonstrate significant improvements over state-of-the-art models for image inpainting, highlighting the efficacy of our approach.

Recognition of Conus species using a combined approach of supervised learning and deep learning-based feature extraction.

Cone snails are venomous marine gastropods comprising more than 950 species widely distributed across different habitats. Their conical shells are remarkably similar to those of other invertebrates in terms of color, pattern, and size. For these reasons, assigning taxonomic signatures to cone snail shells is a challenging task. In this report, we propose an ensemble learning strategy based on the combination of Random Forest (RF) and XGBoost (XGB) methods. We used 47,600 cone shell images of uniform size (224 x 224 pixels), which were split into an 80:20 train-test ratio. Prior to performing subsequent operations, these images were subjected to pre-processing and transformation. After applying a deep learning approach (Visual Geometry Group with a 16-layer deep model architecture) for feature extraction, model specificity was further assessed by including multiple related and unrelated seashell images. Both classifiers demonstrated comparable recognition ability on random test samples. The evaluation results suggested that RF outperformed XGB due to its high accuracy in recognizing Conus species, with an average precision of 95.78%. The area under the receiver operating characteristic curve was 0.99, indicating the model's optimal performance. The learning and validation curves also demonstrated a robust fit, with the training score reaching 1 and the validation score gradually increasing to 95 as more data was provided. These values indicate a well-trained model that generalizes effectively to validation data without significant overfitting. The gradual improvement in the validation score curve is crucial for ensuring model reliability and minimizing the risk of overfitting. Our findings revealed an interactive visualization. The performance of our proposed model suggests its potential for use with datasets of other mollusks, and optimal results may be achieved for their categorization and taxonomical characterization.

Deep Learning-Assisted Design of Novel Donor-Acceptor Combinations for Organic Photovoltaic Materials with Enhanced Efficiency.

Designing donor (D) and acceptor (A) structures and discovering promising D-A combinations can effectively improve organic photovoltaic (OPV) device performance. However, to obtain excellent power conversion efficiency (PCE), the trial-and-error structural design in the infinite chemical space is time-consuming and costly. Herein, a deep learning (DL)-assisted design framework for OPV materials is proposed. To effectively digitally represent the D and A structures, a structure representation method, polymer fingerprints, is developed, and a database of OPV materials is constructed. By applying an end-to-end graph neural network modeling method, high-precision DL models for predicting OPV performance are established. After combining the existing structures, ≈0.6 million virtual D-A combinations are generated. Then, the OPV performance of these candidate combinations is predicted by the well-trained models, and numbers of novel D-A combinations with high efficiency are identified. Experimental validations confirm that the prediction accuracy is greater than 93% and one of the screened combinations (i.e., D18:BTP-S11) exhibits an efficiency above 19.3% in single-junction organic solar cells. Finally, based on the structural gene analysis, the design rules to guide experimental explorations are suggested. The developed DL-assisted approach can accelerate the design of D-A combinations with ultrahigh efficiency and bring property breakthroughs for OPV devices.

FormulationBCS: A Machine Learning Platform Based on Diverse Molecular Representations for Biopharmaceutical Classification System (BCS) Class Prediction.

The Biopharmaceutics Classification System (BCS) has facilitated biowaivers and played a significant role in enhancing drug regulation and development efficiency. However, the productivity of measuring the key discriminative properties of BCS, solubility and permeability, still requires improvement, limiting high-throughput applications of BCS, which is essential for evaluating drug candidate developability and guiding formulation decisions in the early stages of drug development. In recent years, advancements in machine learning (ML) and molecular characterization have revealed the potential of quantitative structure-performance relationships (QSPR) for rapid and accurate in silico BCS classification. The present study aims to develop a web platform for high-throughput BCS classification based on high-performance ML models. Initially, four data sets of BCS-related molecular properties: log S, log P, log D, and log Papp were curated. Subsequently, 6 ML algorithms or deep learning frameworks were employed to construct models, with diverse molecular representations ranging from one-dimensional molecular fingerprints, descriptors, and molecular graphs to three-dimensional molecular spatial coordinates. By comparing different combinations of molecular representations and learning algorithms, LightGBM exhibited excellent performance in solubility prediction, with an R2 of 0.84; AttentiveFP outperformed others in permeability prediction, with R2 values of 0.96 and 0.76 for log P and log D, respectively; and XGBoost was the most accurate for log Papp prediction, with an R2 of 0.71. When externally validated on a marketed drug BCS category data set, the best-performing models achieved classification accuracies of over 77 and 73% for solubility and permeability, respectively. Finally, the well-trained models were embedded into the first ML-based BCS class prediction web platform (x f), enabling pharmaceutical scientists to quickly determine the BCS category of candidate drugs, which will aid in the high-throughput BCS assessment for candidate drugs during the preformulation stage, thereby promoting reduced risk and enhanced efficiency in drug development and regulation.

WTDP-Shapley: Efficient and Effective Incentive Mechanism in Federated Learning for Intelligent Safety Inspection

Artificial Intelligence (AI) has been widely applied for Safety Inspection in a number of industrial domains. However, individual company usually could not provide sufficient data to support well-trained AI models. Federated Learning, as a new AI paradigm, enables a number of participants to contribute training data to co-create high-performance models without compromising data privacy. However, an effective incentive mechanism is essential to encourage participants to contribute high-quality data, and the fundamental of the incentive mechanism is to evaluate participants' contribution fairly. Shapley Value (SV) is a well-known approach to evaluate individual's marginal contribution in a coalition, but the canonical SV calculation and its available variants are very costly. In this paper, we proposed an FL framework to enable a number of natural gas companies from different cities to jointly train an object detection computer vision deep learning model for the purpose of identifying potential hazards, without sharing their confidential inspection photos directly. We improve state-of-the-art SV algorithm by proposing Weighted Truncation (WT) for unnecessary computations, and go further with Dynamic Programming (WTDP), achieving better trade-off between efficiency and accuracy. Based on our proposed WTDP-Shapley participant contribution evaluation approach, an effective end-to-end incentive mechanism is designed by leveraging knowledge of both data scientists and domain experts. According to our experiments, it could encourage participants to contribute scarce photos with potential hazards, and thus co-create a high-performance AI model to identify various hazards accurately for residential natural gas installation safety inspection.

A Curriculum-Style Self-Training Approach for Source-Free Semantic Segmentation.

Source-free domain adaptation has developed rapidly in recent years, where the well-trained source model is adapted to the target domain instead of the source data, offering the potential for privacy concerns and intellectual property protection. However, a number of feature alignment techniques in prior domain adaptation methods are not feasible in this challenging problem setting. Thereby, we resort to probing inherent domain-invariant feature learning and propose a curriculum-style self-training approach for source-free domain adaptive semantic segmentation. In particular, we introduce a curriculum-style entropy minimization method to explore the implicit knowledge from the source model, which fits the trained source model to the target data using certain information from easy-to-hard predictions. We then train the segmentation network by the proposed complementary curriculum-style self-training, which utilizes the negative and positive pseudo labels following the curriculum-learning manner. Although negative pseudo-labels with high uncertainty cannot be identified with the correct labels, they can definitely indicate absent classes. Moreover, we employ an information propagation scheme to further reduce the intra-domain discrepancy within the target domain, which could act as a standard post-processing method for the domain adaptation field. Furthermore, we extend the proposed method to a more challenging black-box source model scenario where only the source model's predictions are available. Extensive experiments validate that our method yields state-of-the-art performance on source-free semantic segmentation tasks for both synthetic-to-real and adverse conditions datasets.

Domain-guided conditional diffusion model for unsupervised domain adaptation.

Limited transferability hinders the performance of a well-trained deep learning model when applied to new application scenarios. Recently, Unsupervised Domain Adaptation (UDA) has achieved significant progress in addressing this issue via learning domain-invariant features. However, the performance of existing UDA methods is constrained by the possibly large domain shift and limited target domain data. To alleviate these issues, we propose a Domain-guided Conditional Diffusion Model (DCDM), which generates high-fidelity target domain samples, making the transfer from source domain to target domain easier. DCDM introduces class information to control labels of the generated samples, and a domain classifier to guide the generated samples towards the target domain. Extensive experiments on various benchmarks demonstrate that DCDM brings a large performance improvement to UDA.

Interactive Prognosis Framework Between Deep Learning and a Stochastic Process Model for Remaining Useful Life Prediction.

Uncertainty quantification of the remaining useful life (RUL) for degraded systems under the big data era has been a hot topic in recent years. A general idea is to execute two separate steps: deep-learning-based health indicator (HI) construction and stochastic process-based degradation modeling. However, there exists a critical matching defect between the constructed HI and a degradation model, which seriously affects the RUL prediction accuracy. Toward this end, this article proposes an interactive prognosis framework between deep learning and a stochastic process model for the RUL prediction. First, we resort to stacked contractive autoencoders to fuse multiple sensor information of historical systems for constructing the HI in a typical unsupervised manner. Then, considering the nonlinear characteristic of the constructed HI, an exponential-like degradation model is introduced to construct its degradation evolving model, and theoretical expressions of the prediction results are derived under the concept of the first hitting time. Furthermore, we design an optimization objective function by integrating the HI construction and degradation modeling for the RUL prediction. To minimize the designed objective function of the proposed interactive prognosis framework, a gradient descent algorithm is employed to update the model parameters. Based on the well-trained interactive prognosis model, we can obtain the HI of a field system from stacked contractive autoencoders with sensor data and the probability density function (pdf) of the predicted RUL on the basis of the estimated parameters. Finally, the effectiveness and superiority of the proposed interactive prognosis method are verified by two case studies associated with turbofan engines.

Say No to Freeloader: Protecting Intellectual Property of Your Deep Model.

Model intellectual property (IP) protection has gained attention due to the significance of safeguarding intellectual labor and computational resources. Ensuring IP safety for trainers and owners is critical, especially when ownership verification and applicability authorization are required. A notable approach involves preventing the transfer of well-trained models from authorized to unauthorized domains. We introduce a novel Compact Un-transferable Pyramid Isolation Domain (CUPI-Domain) which serves as a barrier against illegal transfers from authorized to unauthorized domains. Inspired by human transitive inference, the CUPI-Domain emphasizes distinctive style features of the authorized domain, leading to failure in recognizing irrelevant private style features on unauthorized domains. To this end, we propose CUPI-Domain generators, which select features from both authorized and CUPI-Domain as anchors. These generators fuse the style features and semantic features to create labeled, style-rich CUPI-Domain. Additionally, we design external Domain-Information Memory Banks (DIMB) for storing and updating labeled pyramid features to obtain stable domain class features and domain class-wise style features. Based on the proposed whole method, the novel style and discriminative loss functions are designed to effectively enhance the distinction in style and discriminative features between authorized and unauthorized domains. We offer two solutions for utilizing CUPI-Domain based on whether the unauthorized domain is known: target-specified CUPI-Domain and target-free CUPI-Domain. Comprehensive experiments on various public datasets demonstrate the effectiveness of our CUPI-Domain approach with different backbone models, providing an efficient solution for model intellectual property protection.

Cross-center Model Adaptive Tooth segmentation.

Automatic 3-dimensional tooth segmentation on intraoral scans (IOS) plays a pivotal role in computer-aided orthodontic treatments. In practice, deploying existing well-trained models to different medical centers suffers from two main problems: (1) the data distribution shifts between existing and new centers, which causes significant performance degradation. (2) The data in the existing center(s) is usually not permitted to be shared, and annotating additional data in the new center(s) is time-consuming and expensive, thus making re-training or fine-tuning unfeasible. In this paper, we propose a framework for Cross-center Model Adaptive Tooth segmentation (CMAT) to alleviate these issues. CMAT takes the trained model(s) from the source center(s) as input and adapts them to different target centers, without data transmission or additional annotations. CMAT is applicable to three cross-center scenarios: source-data-free, multi-source-data-free, and test-time. The model adaptation in CMAT is realized by a tooth-level prototype alignment module, a progressive pseudo-labeling transfer module, and a tooth-prior regularized information maximization module. Experiments under three cross-center scenarios on two datasets show that CMAT can consistently surpass existing baselines. The effectiveness is further verified with extensive ablation studies and statistical analysis, demonstrating its applicability for privacy-preserving model adaptive tooth segmentation in real-world digital dentistry.

[formula omitted]-norm distortion-efficient adversarial attack

Adversarial examples have shown a powerful ability to make a well-trained model misclassified. Current mainstream adversarial attack methods only consider one of the distortions among L0-norm, L2-norm, and L∞-norm. L0-norm based methods cause large modification on a single pixel, resulting in naked-eye visible detection, while L2-norm and L∞-norm based methods suffer from weak robustness against adversarial defense since they always diffuse tiny perturbations to all pixels. A more realistic adversarial perturbation should be sparse and imperceptible. In this paper, we propose a novel Lp-norm distortion-efficient adversarial attack, which not only owns the least L2-norm loss but also significantly reduces the L0-norm distortion. To this aim, we design a new optimization scheme, which first optimizes an initial adversarial perturbation under L2-norm constraint, and then constructs a dimension unimportance matrix for the initial perturbation. Such a dimension unimportance matrix can indicate the adversarial unimportance of each dimension of the initial perturbation. Furthermore, we introduce a new concept of adversarial threshold for the dimension unimportance matrix. The dimensions of the initial perturbation whose unimportance is higher than the threshold will be all set to zero, greatly decreasing the L0-norm distortion. Experimental results on three benchmark datasets show that under the same query budget, the adversarial examples generated by our method have lower L0-norm and L2-norm distortion than the state-of-the-art. Especially for the MNIST dataset, our attack reduces 8.1% L2-norm distortion meanwhile remaining 47% pixels unattacked. This demonstrates the superiority of the proposed method over its competitors in terms of adversarial robustness and visual imperceptibility. The code is available at https://github.com/GZHU-DVL/ZhouChao.

Prediction of unknown compressive stress/damage in concrete structure using 3D convolutional neural network-based deep learning of raw electromechanical admittance signals

Deep learning approach promotes automated identification of concrete structural stress and damage using electromechanical impedance/admittance (EMI/EMA) technique, while currently automatic identification is overly depending on the known label with certain stress level, which hinders its real-life application for those unknown loading conditions. To this end, this paper proposed a novel deep learning approach using three-dimensional convolutional neural network (3DCNN)-enhanced EMA technique to detect unknown stress and its corelated damage in concrete structures. In the approach, sub-range raw conductance signatures according to the critical root mean square deviations were selected to construct 3D input of the CNN model, and the unknown stress and its belonged range were rapidly predicted. Unknown single stress level and multiple stress range prediction were investigated via monitoring of a concrete structure subjected to uniaxial compression from initial loading to final failure. Experimental validation results demonstrated that the unknown stress level in a range as wide as 16 MPa could be rapidly predicted with high accuracy and efficiency by using the well-trained 3DCNN model, which outperformed the existed two-dimensional CNN and traditional back-propagation networks. Promising results in this study potentially provided a possible paradigm of the EMA data-driven compressive stress prediction in concrete structures with limited condition measurements.

Data-driven optimization of nose profiles for water entry impact load reduction

The prediction and reduction of impact loads during water entry are critical in various engineering applications, such as ship slamming and seaplane landings. Traditional methods, including theoretical models, computational fluid dynamics (CFD), and experimental approaches, often suffer from limited applicability, lengthy processing times, and high costs. To address these challenges, this study explores the use of Deep Neural Networks (DNNs) for predicting peak impact loads during the oblique water entry of cylinders with different nose profiles. Optimized Latin Hypercube Sampling (OLHS) and Sobol sequences are employed to generate dataset inputs from a 7-dimensional parameter space that governs impact dynamics. These inputs are then fed into OpenFOAM solvers to compute the peak impact accelerations, serving as dataset outputs for training and testing the DNN. Experiments are conducted to validate the numerical method, confirming that the resulting dataset outputs can be considered as ground truth. With Bayesian hyperparameter optimization, the well-trained DNN model demonstrates reliable accuracy in predicting peak axial and normal accelerations. Further, this study integrates the DNN model with the Non-dominated Sorting Genetic Algorithm II (NSGA-II) to optimize the nose profiles, resulting in a maximum weighted reduction of 48% in the combined axial and normal accelerations.

Random transformations to improve mitigation of query-based black-box attacks

This paper proposes methods to upstage the best-known defences against query-based black-box attacks. These benchmark defences incorporate gaussian noise into input data during inference to achieve state-of-the-art performance in protecting image classification models against the most advanced query-based black-box attacks. Even so there is a need to improve upon them; for example, the widely benchmarked Random noise defense (RND) method has demonstrated limited robustness – achieving only 53.5% and 18.1% with a ResNet-50 model on the CIFAR-10 and ImageNet datasets, respectively – against the square attack, which is commonly regarded as the state-of-the-art black-box attack. Therefore, in this work, we propose two alternatives to gaussian noise addition at inference time: random crop-resize and random rotation of the input images. Although these transformations are generally used for data augmentation while training to improve model invariance and generalisation, their protective potential against query-based black-box attacks at inference time is unexplored. Therefore, for the first time, we report that for such well-trained models either of the two transformations can also blunt powerful query-based black-box attacks when used at inference time on three popular datasets. The results show that the proposed randomised transformations outperform RND in terms of robust accuracy against a strong adversary that uses a high budget of 100,000 queries based on expectation over transformation (EOT) of 10, by 0.9% on the CIFAR-10 dataset, 9.4% on the ImageNet dataset and 1.6% on the Tiny ImageNet dataset. Crucially, in two even tougher attack settings, that is, high-confidence adversarial examples and EOT-50 adversary, these transformations are even more effective as the margin of improvement over the benchmarks increases further.

Simplified Knowledge Distillation for Deep Neural Networks Bridging the Performance Gap with a Novel Teacher–Student Architecture

The rapid evolution of deep learning has led to significant achievements in computer vision, primarily driven by complex convolutional neural networks (CNNs). However, the increasing depth and parameter count of these networks often result in overfitting and elevated computational demands. Knowledge distillation (KD) has emerged as a promising technique to address these issues by transferring knowledge from a large, well-trained teacher model to a more compact student model. This paper introduces a novel knowledge distillation method that simplifies the distillation process and narrows the performance gap between teacher and student models without relying on intricate knowledge representations. Our approach leverages a unique teacher network architecture designed to enhance the efficiency and effectiveness of knowledge transfer. Additionally, we introduce a streamlined teacher network architecture that transfers knowledge effectively through a simplified distillation process, enabling the student model to achieve high accuracy with reduced computational demands. Comprehensive experiments conducted on the CIFAR-10 dataset demonstrate that our proposed model achieves superior performance compared to traditional KD methods and established architectures such as ResNet and VGG networks. The proposed method not only maintains high accuracy but also significantly reduces training and validation losses. Key findings highlight the optimal hyperparameter settings (temperature T = 15.0 and smoothing factor α = 0.7), which yield the highest validation accuracy and lowest loss values. This research contributes to the theoretical and practical advancements in knowledge distillation, providing a robust framework for future applications and research in neural network compression and optimization. The simplicity and efficiency of our approach pave the way for more accessible and scalable solutions in deep learning model deployment.

Facilitated the discovery of new γ/γ′ Co-based superalloys by combining first-principles and machine learning

Superalloys are indispensable materials for the fabrication of high-temperature components in aircraft engines. The discovery of a novel class of γ/γ′ Co-Al-W alloys has ignited a surge of interest in Co-based superalloys, with the aspiration to transcend the inherent constraints of their Ni-based counterparts. However, the conventional methodologies utilized in the design and advancement of new γ/γ′ Co-based superalloys are frequently characterized by their laborious and resource-intensive nature. In this study, we employed a coupled Density Functional Theory (DFT) and machine learning (ML) approach to predict and analyze the stability of the crucial γ′ phase, which is instrumental in expediting the discovery of γ/γ′ Co-based alloys. A dataset comprised of thousands of reliable formation (Hf) and decomposition (Hd) energies was obtained through high-throughput DFT calculations. Through regression model selection and feature engineering, our trained Random Forest (RF) model achieved prediction accuracies of 98.07% for Hf and 97.05% for Hd. Utilizing the well-trained RF model, we predicted the energies of over 150,000 ternary and quaternary γ′ phases within the Co-Ni-Fe-Cr-Al-W-Ti-Ta-V-Mo-Nb system. The energy analyses revealed that the presence of Ni, Nb, Ta, Ti, and V significantly reduced the Hf and the Hd of γ′, while Mo and W deteriorate the stability by increasing both energy values. Interestingly, although Al reduces the Hf, it increases Hd, thereby adversely affecting the stability of γ′. Applying domain-specific screening based on our knowledge, we identified 1049 out of >150,000 compositions likely to form stable γ′ phases, predominantly distributed across 11 Al-containing systems and 25 Al-free systems. Combining the analysis of CALPHAD method, we experimentally synthesized two new Co-based alloys with γ/γ′ dual-phase microstructures, corroborating the reliability of our theoretical prediction model.

Recommendation Unlearning via Influence Function

Recommendation unlearning is an emerging task to serve users for erasing unusable data ( e.g., some historical behaviors) from a well-trained recommender model. Existing methods process unlearning requests by fully or partially retraining the model after removing the unusable data. However, these methods are impractical due to the high computation cost of full retraining and the highly possible performance damage of partial training. In this light, a desired recommendation unlearning method should obtain a similar model as full retraining in a more efficient manner, i.e., achieving complete, efficient and harmless unlearning. In this work, we propose a new Influence Function-based Recommendation Unlearning (IFRU) framework, which efficiently updates the model without retraining by estimating the influence of the unusable data on the model via the influence function . In the light that recent recommender models use historical data for both the constructions of the optimization loss and the computational graph ( e.g., neighborhood aggregation), IFRU jointly estimates the direct influence of unusable data on optimization loss and the spillover influence on the computational graph to pursue complete unlearning. Furthermore, we propose an importance-based pruning algorithm to reduce the cost of the influence function. IFRU is harmless and applicable to mainstream differentiable models. Extensive experiments demonstrate that IFRU achieves more than 250 times acceleration compared to retraining-based methods with recommendation performance comparable to full retraining. Codes are available at https://github.com/baiyimeng/IFRU.

SSAT: Active Authorization Control and User’s Fingerprint Tracking Framework for DNN IP Protection

As training a high-performance deep neural network (DNN) model requires a large amount of data, powerful computing resources and expert knowledge, protecting well-trained DNN models from intellectual property (IP) infringement has raised serious concerns in recent years. Most existing methods using DNN watermarks to verify the ownership of the models after IP infringement occurs, which is reactive in the sense that they cannot prevent unauthorized users from using the model in the first place. Different from these methods, in this article, we propose an active authorization control and user’s fingerprint tracking method for the IP protection of DNN models by utilizing sample-specific backdoor attack. The proposed method inversely and multiplely exploits sample-specific trigger as the key to implement authorization control for DNN model, in which the generated triggers are imperceptible and sample-specific for clean images. Specifically, a U-Net model is used to generate backdoor instances. Then, the target model is trained on the clean images and backdoor instances, which are inversely labeled as wrong classes and correct classes, respectively. Only authorized users can use the target model normally by pre-processing the clean images through the U-Net model. Moreover, the images processed by the U-Net model will contain unique fingerprint that can be extracted to verify and track the corresponding user’s identity. This article is the first work that utilizes the sample-specific backdoor attack to implement active authorization control and user’s fingerprint management for DNN model under black-box scenarios. Extensive experimental results on ImageNet dataset and YouTube Aligned Face dataset demonstrate that the proposed method is effective in protecting the DNN model from unauthorized usage. Specifically, the protected model has a low inference accuracy (1.00%) for unauthorized users, while maintaining a normal inference accuracy (97.67%) for authorized users. Besides, the proposed method can achieve 100% fingerprint tracking success rates on both the ImageNet and YouTube Aligned Face datasets. Moreover, it is demonstrated that the proposed method is robust against fine-tuning attack, pruning attack, pruning attack with retraining, reverse-engineering attack, adaptive attack, and JPEG compression attack. The code is available at https://github.com/nuaaaisec/SSAT .