CHI HONG AO

Followers

Following

Public Views

Interests

Uploads

Papers by CHI HONG AO

AGIC: Approximate Gradient Inversion Attack on Federated Learning

2022 41st International Symposium on Reliable Distributed Systems (SRDS)

Federated learning is a private-by-design distributed learning paradigm where clients train local... more Federated learning is a private-by-design distributed learning paradigm where clients train local models on their own data before a central server aggregates their local updates to compute a global model. Depending on the aggregation method used, the local updates are either the gradients or the weights of local learning models, e.g., FedAvg aggregates model weights. Unfortunately, recent reconstruction attacks apply a gradient inversion optimization on the gradient update of a single minibatch to reconstruct the private data used by clients during training. As the state-of-the-art reconstruction attacks solely focus on single update, realistic adversarial scenarios are overlooked, such as observation across multiple updates and updates trained from multiple mini-batches. A few studies consider a more challenging adversarial scenario where only model updates based on multiple mini-batches are observable, and resort to computationally expensive simulation to untangle the underlying samples for each local step. In this paper, we propose AGIC, a novel Approximate Gradient Inversion Attack that efficiently and effectively reconstructs images from both model or gradient updates, and across multiple epochs. In a nutshell, AGIC (i) approximates gradient updates of used training samples from model updates to avoid costly simulation procedures, (ii) leverages gradient/model updates collected from multiple epochs, and (iii) assigns increasing weights to layers with respect to the neural network structure for reconstruction quality. We extensively evaluate AGIC on three datasets, namely CIFAR-10, CIFAR-100 and ImageNet. Our results show that AGIC increases the peak signal-to-noise ratio (PSNR) by up to 50% compared to two representative state-of-the-art gradient inversion attacks. Furthermore, AGIC is faster than the state-of-the-art simulationbased attack, e.g., it is 5x faster when attacking FedAvg with 8 local steps in between model updates. Index Terms-reconstruction attack, federated learning, federated averaging

Download

The Relationships between Metacognition and Reading Strategies: The Case of English-Majored Students at a Private University in Vietnam

European Journal of Educational Research

Many studies have pointed out a proportionate relation between metacognition and reading strategi... more Many studies have pointed out a proportionate relation between metacognition and reading strategies. This study advances such an understanding by arguing that this relation is comprised of several sub-relations. Metacognitive Awareness Inventory and Reading Strategies Inventory were conducted among 92 English as Foreign Language students at a Vietnamese university. This study highlights 3 major findings. First, metacognition processes do not merely mean “cognition about cognition”. Second, while previous studies tended to portray metacognition as a whole, it is argued in this study that it is constituted by two main clusters (knowledge of cognition and regulation of cognition) which can be further specified into eight sub-components. Third, there seem to be some sub-linearities that underpin students’ metacognition and that influence their uses of reading strategies. These findings are hoped to shed light on preserving spaces and autonomy in curriculum/syllabus development for stude...

Download

MEGA: Model Stealing via Collaborative Generator-Substitute Networks

arXiv (Cornell University), Jan 31, 2022

Deep machine learning models are increasingly deployed in the wild for providing services to user... more Deep machine learning models are increasingly deployed in the wild for providing services to users. Adversaries may steal the knowledge of these valuable models by training substitute models according to the inference results of the targeted deployed models. Recent data-free model stealing methods are shown effective to extract the knowledge of the target model without using real query examples, but they assume rich inference information, e.g., class probabilities and logits. However, they are all based on competing generatorsubstitute networks and hence encounter training instability. In this paper we propose a data-free model stealing framework, MEGA, which is based on collaborative generatorsubstitute networks and only requires the target model to provide label prediction for synthetic query examples. The core of our method is a model stealing optimization consisting of two collaborative models (i) the substitute model which imitates the target model through the synthetic query examples and their inferred labels and (ii) the generator which synthesizes images such that the confidence of the substitute model over each query example is maximized. We propose a novel coordinate descent training procedure and analyze its convergence. We also empirically evaluate the trained substitute model on three datasets and its application on black-box adversarial attacks. Our results show that the accuracy of our trained substitute model and the adversarial attack success rate over it can be up to 33% and 40% higher than state-of-the-art data-free black-box attacks.

Download

Is Shapley Value fair? Improving Client Selection for Mavericks in Federated Learning

arXiv (Cornell University), Jun 20, 2021

Shapley Value is commonly adopted to measure and incentivize client participation in federated le... more Shapley Value is commonly adopted to measure and incentivize client participation in federated learning. In this paper, we show-theoretically and through simulations-that Shapley Value underestimates the contribution of a common type of client: the Maverick. Mavericks are clients that differ both in data distribution and data quantity and can be the sole owners of certain types of data. Selecting the right clients at the right moment is important for federated learning to reduce convergence times and improve accuracy. We propose FEDEMD, an adaptive client selection strategy based on the Wasserstein distance between the local and global data distributions. As FEDEMD adapts the selection probability such that Mavericks are preferably selected when the model benefits from improvement on rare classes, it consistently ensures the fast convergence in the presence of different types of Mavericks. Compared to existing strategies, including Shapley Value-based ones, FEDEMD improves the convergence of neural network classifiers by at least 26.9% for FedAvg aggregation compared with the state of the art. Preprint. Under review.

Download

End-to-End Learning from Noisy Crowd to Supervised Machine Learning Models

2020 IEEE Second International Conference on Cognitive Machine Intelligence (CogMI), 2020

Labeling real-world datasets is time consuming but indispensable for supervised machine learning ... more Labeling real-world datasets is time consuming but indispensable for supervised machine learning models. A common solution is to distribute the labeling task across a large number of non-expert workers via crowd-sourcing. Due to the varying background and experience of crowd workers, the obtained labels are highly prone to errors and even detrimental to the learning models. In this paper, we advocate using hybrid intelligence, i.e., combining deep models and human experts, to design an end-to-end learning framework from noisy crowd-sourced data, especially in an on-line scenario. We first summarize the state-ofthe-art solutions that address the challenges of noisy labels from non-expert crowd and learn from multiple annotators. We show how label aggregation can benefit from estimating the annotators' confusion matrices to improve the learning process. Moreover, with the help of an expert labeler as well as classifiers, we cleanse aggregated labels of highly informative samples to enhance the final classification accuracy. We demonstrate the effectiveness of our strategies on several image datasets, i.e. UCI and CIFAR-10, using SVM and deep neural networks. Our evaluation shows that our on-line label aggregation with confusion matrix estimation reduces the error rate of labels by over 30%. Furthermore, relabeling only 10% of the data using the expert's results in over 90% classification accuracy with SVM.

Download

PASCMP: A novel cache framework for data mining application

2016 2nd IEEE International Conference on Computer and Communications (ICCC), 2016

In Chip Multiprocessors (CMP) architecture, traditional shared last level cache (LLC) is widely u... more In Chip Multiprocessors (CMP) architecture, traditional shared last level cache (LLC) is widely used. Shared LLC has a large aggregate cache capacity, so it able to minimizes off-chip memory requests. However, with the number of cores increasing, the global wire delay affects cache performance greatly. Data mining is the main method of processing data and knowledge discovery. By analyzing the characterizations of some data mining programs, we find out that the proportion of shared data are relatively less, and a significant portion of accesses are related to private data. Shared LLC does not fit data mining applications,because private data may be placed in a LLC bank that far away from the core which requests the data. In order to adapt to this kind of programs, our new architecture PASCMP makes a trade-off between capacity and wire delay. We design placement and migration strategies to respectively put private data and shared data in the right place at a minimum cost. Thus each co...

Timeliness assessment function construction and property analysis of emergency response procedures

Multi-agent cooperation based on organizational structure

Computer Engineering, 2005

Agent organization can be used for multi-agent problem solving so as to potentially decrease the ... more Agent organization can be used for multi-agent problem solving so as to potentially decrease the difficulty of problem solving and complexity of interaction, and the organization structure, organization formation rule and object planning are the key problems among multi-agent cooperation. In this paper, a multi-agent organization model and a set of rules based on the model were presented and the method about multi-agent object planning and the coherence and perfection of the planning were described. Based on the above study, the formation procedure and the release procedure of multi-agent organization were presented, and all the procedures were formally described by extended temporal logic and pi-calculus. All these improved the study of multi-agent theory and programming practice which was not closely combined.

Study on the Relationship Between Abnormal Behavior and Blood Lead Level Among the Children in Nanjing

Chinese Journal of School Health, 2006

Objective To explore the relationship between the abnormal behavior and blood lead level of the c... more Objective To explore the relationship between the abnormal behavior and blood lead level of the children,and to provide bases for the prevention measures.Methods A total of 404 children aged 9-12 in two primary schools in Nanjing were selected with random cluster sampling.40 μL blood from the finger was taken for all the children,then digested and centrifuged,the lead concentration in the whole blood was measured with graphite furnace atom absorption spectrometry.In the mean time,children's behavior was measured with Achenbach Behavior Checklist.Results The average blood lead level for the children was 81.2 μg/L.The blood lead level and the number of blood lead poisoning for the boys were more than those for girls,and the blood lead level and the prevalence of lead poisoning for children with abnormal behavior were obviously more than those for children with normal behavior.Conclusion Risk of lead poisoning is serious for boys than for girls.There is relationship between blood l...

A Study of SS Series Additives for Acidic Bright Tin Plating

Electroplating & Pollution Control, 2002

The effects of main auxiliary brighteners, formaldehyde dispersing agent and stabilizer on bath p... more

Label Aggregation via Finding Consensus Between Models

ArXiv, 2018

Label aggregation is an efficient and low cost way to make large datasets for supervised learning... more Label aggregation is an efficient and low cost way to make large datasets for supervised learning. It takes the noisy labels provided by non-experts and infers the unknown true labels. In this paper, we propose a novel label aggregation algorithm which includes a label aggregation neural network. The learning task in this paper is unsupervised. In order to train the neural network, we try to design a suitable guiding model to define the loss function. The optimization goal of our algorithm is to find the consensus between the predictions of the neural network and the guiding model. This algorithm is easy to optimize using mini-batch stochastic optimization methods. Since the choices of the neural network and the guiding model are very flexible, our label aggregation algorithm is easy to extend. According to the algorithm framework, we design two novel models to aggregate noisy labels. Experimental results show that our models achieve better results than state-of-the-art label aggreg...

Download

Verification of“ Three Dates and One Certificate”in Personnel Files

Journal of Anhui Agricultural University, 2004

Employees'“ three dates and one certificate (date of birth,date of joining Chinese Communist ... more Employees'“ three dates and one certificate (date of birth,date of joining Chinese Communist Party,date of first employment,and certificate of highest education degree)” are the basis of human resource information management as well as important references for further decision- making processes.Verification of the“ three dates and one certificate” by the human resource department is a very serious work which requires great care and strict methods of relative policies.It influences directly on each employee's individual interest,and also has effect on the quality and efficiency of the human resource management work. But recently for some employees,the records of the“ three dates and one certificate” are not consistent with fact or historical ones,for such problems,the author proposes some detailed verification methods to distinguish the true ones. These methods not only aim to safeguard individual employee's personal interest, but also try to eliminate the“ alias” problem...

Study on the relationship between the abnormal behavior and blood lead level of the children in Nanjing

Jiangsu Journal of Preventive Medicine, 2006

Objective: To learn the relationship between the abnormal behavior and blood lead level in childr... more Objective: To learn the relationship between the abnormal behavior and blood lead level in children.Methods: Four hundred and four children aged 9-12 years old at 2 primary schools in Nanjing were selected with random cluster sampling.Fortyμl blood was taken from finger for every child and the lead concentration in the whole blood was measured by graphite furnance atom absorption spectrometry after being digested and centrifuge with all blood samples.The children whose blood lead level was over 100μg/L were divided into the lead poisoning group.In the mean time,children's behavior was measured by Achenbach behavior checklist,which filled by children's parents.Results: It was found that blood lead level in average for the children was 81.2μg/L;The blood lead level and the numbers of lead poisoning for the boys were more than those for girls,and the blood lead level and the prevalence of lead poisoning for children with abnormal behavior were obviously higher than those for ch...

Prevalence, risk factors, and trajectories of sleep disturbance in a cohort of African-American breast cancer survivors

Supportive Care in Cancer, 2020

Purpose Sleep disturbance may be an overlooked modifiable risk factor for health disparities amon... more Purpose Sleep disturbance may be an overlooked modifiable risk factor for health disparities among African-American breast cancer survivors (AABCS). This study aimed to identify the prevalence of and risk factors for sleep disturbance in a cohort of AABCS. Methods The study was conducted among participants in the Women's Circle of Health Follow-up Study, a longitudinal study of breast cancer in 10 counties in New Jersey. Cases were identified shortly after diagnosis by the New Jersey State Cancer Registry. Self-reported sleep disturbance (Pittsburgh Sleep Quality Index) and other factors (e.g., socioeconomic status, menopausal status) were assessed at pre-diagnosis (n = 637), 10 months post-diagnosis (n = 261), and 24 months post-diagnosis (n = 632). Clinical data were obtained via medical record abstraction, and height and weight were measured by study staff. Results Most AABCS (57%) reported clinically significant sleep disturbance before diagnosis, and this rate remained largely unchanged at 10 months (53%) and 24 months post-diagnosis (61%). Average sleep disturbance scores indicated clinically significant disturbance at all three assessments (M range = 6.67-7.57). Most reported sleeping fewer than the recommended 7 hours per night at each assessment (range 57-65%). Risk factors for sleep disturbance were identified at each assessment, including pre-diagnosis (less education), 10 months post-diagnosis (lack of insurance, treatment with chemotherapy), and 24 months post-diagnosis (younger age, less education, lower income, obesity, and lymphedema). Treatment with endocrine therapy was a protective factor at 10 months post-diagnosis. Conclusion Most AABCS report clinically significant sleep disturbance from before diagnosis through 24 months post-diagnosis. These rates appear indicate AABCS experience significant sleep-related disparities.

Download

A survey of microRNA single nucleotide polymorphisms identifies novel breast cancer susceptibility loci in a case-control, population-based study of African-American women

Breast cancer research : BCR, Jan 5, 2018

MicroRNAs (miRNAs) regulate gene expression and influence cancer. Primary transcripts of miRNAs (... more MicroRNAs (miRNAs) regulate gene expression and influence cancer. Primary transcripts of miRNAs (pri-miRNAs) are poorly annotated and little is known about the role of germline variation in miRNA genes and breast cancer (BC). We sought to identify germline miRNA variants associated with BC risk and tumor subtype among African-American (AA) women. Under the African American Breast Cancer Epidemiology and Risk (AMBER) Consortium, genotyping and imputed data from four studies on BC in AA women were combined into a final dataset containing 224,188 miRNA gene single nucleotide polymorphisms (SNPs) for 8350 women: 3663 cases and 4687 controls. The primary miRNA sequence was identified for 566 miRNA genes expressed in Encyclopedia of DNA Elements (ENCODE) Tier 1 cell types and human pancreatic islets. Association analysis was conducted using logistic regression for BC status overall and by tumor subtype. A novel BC signal was localized to an 8.6-kb region of 17q25.3 by four SNPs (rs9913477...

Download

Role of the meridional dipole of SSTA and associated cross-equatorial flow in the tropical eastern Pacific in terminating the 2014 El Niño development

Climate Dynamics, 2017

In the boreal spring of 2014, the oceanic and atmospheric conditions were favorable for an El Niñ... more In the boreal spring of 2014, the oceanic and atmospheric conditions were favorable for an El Niño's development. It was predicted that in 2014, a super El Niño or at least a regular El Niño with normal magnitude, would initiate. However, the growth rate of the sea surface temperature anomaly (SSTA) in the equatorial eastern Pacific suddenly declined in the boreal summer. The physical processes responsible for the termination of the 2014 El Niño were addressed in this study. We hypothesized that a meridional dipole of SSTA, characterized by a pronounced warm SSTA over the eastern North Pacific (ENP) and cold SSTA over the eastern South Pacific (ESP), played a crucial role in blocking the 2014 El Niño’s development. The observational analysis revealed that the meridional dipole of SSTA and the relevant anomalous cross-equatorial flow in the tropical eastern Pacific, induced anomalous westward ($${u^\prime }<0$$u′<0) and upwelling ($${w^\prime }>0$$w′>0) currents in the equatorial eastern Pacific, leading to negative anomalous zonal advection term ($$- {u^\prime }\partial \overline T /\partial x<0$$-u′∂T¯/∂x<0) and anomalous upwelling advection term ($$- {w^\prime }\partial \overline T /\partial z<0$$-w′∂T¯/∂z<0). Additionally, the anomalous cross-equatorial flow also induced northward meridional current anomalies that transported subtropical cold water to the equator. All the changes of the oceanic dynamic terms collectively caused negative SSTA tendency in the boreal summer, and thus killed off the budding 2014 El Niño. The idealized numerical experiments further confirmed that the 2014 El Niño’s development could be suppressed by the meridional dipole of SSTA, and both the ENP pole and ESP pole make a contribution.

A Study on the Seismic Retrofit of Column in Educational Facilities Using Composite Material

The Journal of Korean Institute of Educational Facilities, 2013

In paper after the strong earthquake of recently the Korea neighborhood, the Korean government su... more In paper after the strong earthquake of recently the Korea neighborhood, the Korean government survey show that the 86% of school buildings in Korea are in potential damage risk and only 14% of them are designed as earthquake-resistance buildings. Reinforcing projects of school have been conducting by the ministry of education, however their reinforcing methods done by not proved a engineering by experiment which results in uneconomical and uneffective rehabilitation for the future earthquake. An experimental and analytical study have been conducted for the shear and flexural reinforcing method of RC beam using composite beam. Based on the previous research, in this study, performance evaluation for the column reinforcing of old school buildings using nonlinear analysis is going to be conducted and strengthening method is going to be on the market after their performance is proved by the test.

Download

Biology and Etiology of Young-Onset Breast Cancers among Premenopausal African American Women: Results from the AMBER Consortium

Cancer Epidemiology, Biomarkers & Prevention, 2017

Background: African American (AA) women have higher incidence of aggressive, young-onset (<40 ... more Background: African American (AA) women have higher incidence of aggressive, young-onset (<40 years) breast cancers. Young- and older-onset disease may have distinct tumor biologies and etiologies; however, studies investigating age differences among AA women have been rare and generally underpowered.Methods: We examined tumor characteristics and breast cancer risk factors associated with premenopausal young (<40) vs. older (≥40) AA women's breast cancer in the African American Breast Cancer Epidemiology and Risk Consortium (2,008 cases and 5,144 controls). Unconditional logistic regression models assessed heterogeneity of tumor biology and risk factor associations by age, overall, and by estrogen receptor status.Results: Premenopausal AA women <40 years had higher frequency of poorer-prognosis tumor characteristics compared with older women, including negative estrogen and progesterone receptor status, triple-negative subtype, higher grade, higher stage, and larger tum...

Download

A functionally significant SNP in TP53 and breast cancer risk in African-American women

NPJ breast cancer, 2017

A coding region polymorphism exists in the TP53 gene (Pro47Ser; rs1800371) in individuals of Afri... more A coding region polymorphism exists in the TP53 gene (Pro47Ser; rs1800371) in individuals of African descent, which reduces p53 tumor suppressor function in a mouse model. It has been unclear whether this functionally significant polymorphism alters cancer risk in humans. This analysis included 6907 women with breast cancer and 7644 controls from the AMBER, ROOT, and AABC consortia. We used multivariable logistic regression to estimate associations between the TP53 Pro47Ser allele and overall breast cancer risk. Because polymorphisms in TP53 tend to be associated with cancer risk in pre-menopausal women, we also limited our analyses to this population in the AMBER and ROOT consortia, where menopausal status was known, and conducted a fixed effects meta-analysis. In an analysis of all women in the AMBER, ROOT, and AABC consortia, we found no evidence for association of the Pro47Ser variant with breast cancer risk. However, when we restricted our analysis to only pre-menopausal women ...

Download

Genetic variations in the Hippo signaling pathway and breast cancer risk in African American women in the AMBER Consortium

Carcinogenesis, 2016

The Hippo signaling pathway regulates cellular proliferation and survival, thus exerting profound... more The Hippo signaling pathway regulates cellular proliferation and survival, thus exerting profound effects on normal cell fate and tumorigenesis. Dysfunction of the Hippo pathway components has been linked with breast cancer stem cell regulation, as well as breast tumor progression and metastasis. TAZ, a key component of the Hippo pathway, is highly expressed in triple negative breast cancer; however, the associations of genetic variations in this important pathway with breast cancer risk remain largely unexplored. Here, we analyzed 8309 germline variants in 15 genes from the Hippo pathway with a total of 3663 cases and 4687 controls from the African American Breast Cancer Epidemiology and Risk Consortium. Odds ratios (ORs) were estimated using logistic regression for overall breast cancer, by estrogen receptor (ER) status (1983 ER positive and 1098 ER negative), and for case-only analyses by ER status. The Hippo signaling pathway was significantly associated with ER-negative breast cancer (pathway level P = 0.02). Gene-based analyses revealed that CDH1 was responsible for the pathway association (P < 0.01), with rs4783673 in CDH1 statistically significant after genelevel adjustment for multiple comparisons (P = 9.2 × 10 −5 , corrected P = 0.02). rs142697907 in PTPN14 was associated with ER-positive breast cancer and rs2456773 in CDK1 with ER-negativity in case-only analysis after gene-level correction for multiple comparisons (corrected P < 0.05). In conclusion, common genetic variations in the Hippo signaling pathway may contribute to both ER-negative and ER+ breast cancer risk in AA women.

Download

CHI HONG AO

Uploads

Papers by CHI HONG AO

Log In