Current studies have discovered that model extraction attacks (MEA) can steal the functionality of deep learning (DL) models, thus causing economic loss and other security threats. Extraction attackers can build a clone model locally that has a different structure but similar functionality to the victim model. To counter MEA, defenders utilize realistic auxiliary data to enhance the victim model and produce misleading predictions for attack data. However, these defense methods have three critical problems caused by utilizing realistic auxiliary data. First, in some scenarios, realistic auxiliary data is absent and difficult to obtain. Secondly, the defense effectiveness brought by realistic auxiliary data is unstable. Finally, the realistic auxiliary data did not protect all categories of training data, resulting in higher clone accuracy for some categories. To address these issues, we propose Model Defense Variational Autoencoder (MDV) to generate virtual auxiliary data as a replacement for realistic auxiliary data. MDV combines the Variational Autoencoder (VAE) and classifier, compelling the latent features to obey different multivariate Gaussian distributions according to the categories. Then, MDV samples deep features from low-likelihood regions of different distributions as realistic auxiliary data. During the experimental phase, we apply our auxiliary data to different defense methods that use auxiliary data and compare the defense effects in different scenarios. Experimental results demonstrate that our method effectively addressed the three aforementioned issues.

Pre-training techniques based on functional connectivity (FC) have demonstrated great potential in brain disease diagnosis. However, previous studies may disrupt the functional information of training data when designing pre-training tasks, and may be limited by biases stemming from single-task learning or insufficient coordination among multiple task components, thereby hindering the acquisition of robust and generalizable feature representations. To address these limitations, we proposed a novel pre-training framework, named Multi-view Contrastive and Masked ROI Learning (CoMa), to learn general representations from healthy datasets through improved learning tasks, with flexible domain-adaptive fine-tuning for downstream tasks. Results showed that the proposed CoMa achieved superior performance across a broad spectrum of diagnostic tasks, significantly outperforming the alternative methods, emphasizing its generalization and effectiveness. Furthermore, the model can further enhance the diagnostic accuracy through task-specific fine-tuning within particular disease domains, indicating its potential for adaptive disease diagnosis. Additionally, we also identified interpretable diagnostic biomarkers for childhood developmental disorders, psychiatric disorders, and neurodegenerative disorders. Overall, the proposed CoMa is instrumental toward the application of fundamental model for disease diagnosis and improves our understanding of underlying mechanisms of common brain disorders.

Multimodal fusion provides cross-modality information to understand the human brain from different perspectives that may be missed in single modality analysis. Supervised fusion focuses on extracting multimodal patterns related to specific clinical measures by further incorporating a prior interested reference. However, existing supervised fusion methods cannot extract component that have weak correlations with the reference, which may be lost during the optimization process. Here, we propose a reference-guided parallel independent component analysis (RG-PICA) aiming at identifying multimodal covarying features related to interested reference through global optimization. The intra-modality independence, the inter-modality correlation, and the correlation between modalities and the reference are maximized globally. Simulations show that RG-PICA can accurately extract multimodal features correlated with the weak related reference while keeping cross-modality linkage comparing with seven fusion methods. In real data application, RG-PICA reveals co-varying patterns in schizophrenia (SZ) that links with cognition and correlates between modalities. These results demonstrate RG-PICA can jointly optimize for target components that correlate with the reference while keeping cross-modality linkage. This approach can improve the meaningful detection of reliable reference-linked multimodal brain patterns for brain disorders.

As a highly interconnected complex network system, the brain exhibits changes in interactions due to common brain disorders. Studying changes in brain network interactions can help us quantitatively analyze functional network patterns and changes in these patterns that are linked to brain disorders. However, relationships between brain regions estimated by most current approaches use a single connectivity that does not fully reflect multiple interactions. Here, we propose a novel functional connectivity (FC) construction method, which can estimate both cooperative and competitive (C-C) relationships between the same regions of interest (ROIs) through improved Ising model. We redefine the Ising dynamic equation to represent pairwise interactions from single to C-C relationships. Results show that the estimated C-C connectivities are normally distributed, with intra-subjects’ (n=970) similarity being consistently and significantly higher than inter-subjects’ similarity across datasets. C-C FCs between occipital, parietal, temporal cortex and the limbic system of schizophrenia (SZ, n=178) are more competitive, while healthy control (HC, n=219) tends to be more cooperative. Group differences in C-C patterns between SZ and HC show significant differences in frontal, parietal and occipital regions. The proposed C-C approach provide new insights into the brain dysfunction in SZ, which can also be applied to investigate other brain disorders.

Multimodal fusion provides multiple benefits over single modality analysis by leveraging both shared and complementary information from different modalities. Notably, supervised fusion enjoys extensive interest for capturing multimodal co-varying patterns associated with clinical measures. A key challenge of brain data analysis is how to handle confounds, which, if unaddressed, can lead to an unrealistic description of the relationship between the brain and clinical measures. Current approaches often rely on linear regression to remove covariate effects prior to fusion, which may lead to information loss, rather than pursue the more global strategy of optimizing both fusion and covariates removal simultaneously. Thus, we propose "CR-mCCAR" to jointly optimize for confounds within a guided fusion model, capturing co-varying multimodal patterns associated with a specific clinical domain while also discounting covariate effects. Simulations show that CR-mCCAR separate the reference and covariate factors accurately. Functional and structural neuroimaging data fusion reveals co-varying patterns in attention deficit/hyperactivity disorder (ADHD, striato-thalamo-cortical and salience areas) and in autism spectrum disorder (ASD, salience and fronto-temporal areas) that link with core symptoms but uncorrelate with age and motion. These results replicate in an independent cohort. Downstream classification accuracy between ADHD/ASD and controls is markedly higher for CR-mCCAR compared to fusion and regression separately. CR-mCCAR can be extended to include multiple targets and multiple covariates. Overall, results demonstrate CR-mCCAR can jointly optimize for target components that correlate with the reference(s) while removing nuisance covariates. This approach can improve the meaningful detection of reliable phenotype-linked multimodal biomarkers for brain disorders.

Functional magnetic resonance imaging (fMRI), as a non-invasive method to reveal brain function alterations, frequently yields time series with unequal lengths in real-world scenarios, which may arise from factors such as motion artifacts, participant state, and differing scan protocols. This variability conflicts with the traditional methods relying on isometric inputs, which poses a significant challenge for the downstream applications such as brain age prediction. To address this challenge, we introduced Gaussian Process Regression (GPR) to normalize the length of time series and proposed split-channel residual convolution (SC) and self-attention mechanisms (SA) to perform brain age estimation, called GPR-SCSANet. Results showed that the proposed framework, GPR-SCSANet, is able to fully utilize the inherent information and learn richer feature representations from unequal-length fMRI time courses, which significantly improved the prediction accuracy across 3 brain atlases and 5 prediction models. The results demonstrated the effectiveness and robustness of the proposed GPR-SCSANet, showcasing the potential for broader applications in brain age prediction task.

Model extraction (ME) attacks, aiming to steal the functionality or parameters of the victim model, have become a widespread research topic. Most functional ME attack methodologies follow a uniform framework, which we summarize in three steps: initially choosing appropriate attack data, then querying the victim model with this data, and finally, training an incipient clone model based on the victim model’s outputs. Despite much focus on data selection, the latter two steps have been somewhat neglected. Noticing this, we explore a method for the information of attack data labels to enhance the accuracy of the clone model. Specifically, we utilized the incipient clone model to identify similarities between the leaked private data and the attack data, subsequently appending the labels from the leaked data to those of the attack data. Then, we employed these modified attack data labels to fine-tune the incipient clone model, obtaining an enhanced clone model with higher accuracy. The enhancement was applied to three representative ME attack methodologies that primarily focus on the first step. Results show that the enhanced model reveals a higher accuracy than the three basic attacks. In summary, our approach suggests that future research should extend beyond data selection.