Computer Security - ESORICS 2023 : 28th European Symposium on Research in Computer Security, the Hague, the Netherlands, September 25-29, 2023, Proceedings, Part IV / edited by Gene Tsudik [and three others].

Format:

Book

Contributor:

Tsudik, G. (Gene), editor.

Series:

Lecture notes in computer science ; Volume 14347.

Lecture Notes in Computer Science Series ; Volume 14347

Language:

English

Subjects (All):

Computer security--Congresses.

Computer security.

Physical Description:

1 online resource (444 pages)

Edition:

First edition.

Place of Publication:

Cham, Switzerland : Springer, [2024]

Summary:

The four-volume set LNCS 14344-14347 constitutes the refereed proceedings of the 28th European Symposium on Research in Computer Security, ESORICS 2023, which took place in The Hague, The Netherlands, during September 25-29, 2023.The 93 full papers presented in these proceedings were carefully reviewed and selected from 478 submissions.

Contents:

Intro

Preface

Organization

Contents - Part IV

Machine Learning

Reinforcement Learning Approach to Generate Zero-Dynamics Attacks on Control Systems Without State Space Models

1 Introduction

2 Related Works

3 Preliminaries

3.1 Quadruple Tank Process

3.2 State Space-Based Theoretical Attack

3.3 Reinforcement Learning (RL)

3.4 Deep Reinforcement Learning

4 Attacking the Quadruple Tank Process (QTP)

4.1 Description of the QTP Model

4.2 Deep Reinforcement Learning Based Attacker

4.3 Chi-Square Based Anomaly Detector (^2)

4.4 LSTM Based Detector

4.5 Two Proportion Z-Test

5 Results

5.1 RL Based Attackers

5.2 LSTM Based Detectors

5.3 Test of Convergence

5.4 RL vs. Theoretical Attack

6 Conclusion

A Proof for Attack Generation

B Architectures and Hyperparameters

C Attacker Rewards and Performance

References

Secure Split Learning Against Property Inference, Data Reconstruction, and Feature Space Hijacking Attacks

2 Problem Statement

2.1 Deep Learning

2.2 Split Learning

2.3 Threat Model

3 Privacy-Preserving Split Learning

3.1 R3eLU: Randomized-Response ReLU

3.2 Forward Propagation with R3eLU

3.3 Private Backward Propagation

3.4 Dynamic Privacy Budget Allocation

4 Privacy Analysis

5 Evaluation

5.1 Model Usability

5.2 Privacy Preservation

A Appendix

A.1 Model Architecture

A.2 Supplement Results of Evaluation

The Power of MEME: Adversarial Malware Creation with Model-Based Reinforcement Learning

2 Threat Model

3 Background and Related Work

3.1 Reinforcement Learning

3.2 Malware Evasion

3.3 Model Extraction

4 Methodology

4.1 Evasion Evaluation Metrics

4.2 Surrogate Evaluation Metrics

5 Experimental Setup

5.1 Targets

5.2 Datasets.

5.3 Adversarial Malware Generation Comparison

5.4 General Experiment Settings

6 Results

6.1 Malware Evasion

6.2 Surrogate Evaluation

7 Discussion

7.1 Performance Considerations

7.2 The Advantage of Learning a Policy

7.3 Future Work

8 Conclusions

FLGuard: Byzantine-Robust Federated Learning via Ensemble of Contrastive Models

2 Background

2.1 Federated Learning

2.2 Poisoning Attack

2.3 Byzantine-Robust Aggregation Rules

2.4 Contrastive Learning and Clustering

3 Threat Model and Problem Setup

4 Our FLGuard

4.1 Overview

4.2 Preprocessing Local Updates

4.3 Training Contrastive Models

4.4 Filtering Malicious Clients

5.1 Experimental Setup

5.2 Evaluated Poisoning Attacks

5.3 Experimental Result

5.4 Ablation Study

6 Conclusion and Future Work

A.1 FLGuard Algorithms and Parameter Settings for Experiments

A.2 Details of Evaluated Attacks

Machine Learning for SAST: A Lightweight and Adaptable Approach

3 Our Approach

3.1 Embedding Method

3.2 Model Generation

4 Evaluation

5 Discussion

6 Related Research

7 Limitations and Threats to Validity

8 Further Research

9 Conclusion

Two Models are Better Than One: Federated Learning is Not Private for Google GBoard Next Word Prediction

2 Preliminaries

2.1 Federated Learning Client Update

2.2 Threat Model

2.3 GBoard NWP Model

3 Reconstruction Attack

3.1 Word Recovery

3.2 Reconstructing Sentences

4 Performance of Attacks Against Vanilla Federated Learning

4.1 Experimental Setup

4.2 Metrics

4.3 Measurements

5 Existing Attacks and Their Defences

5.1 Image Data Reconstruction

5.2 Text Data Reconstruction

5.3 Proposed Defences.

6 Performance of Our Attacks Against Federated Learning with Local DP

6.1 Word Recovery Performance

6.2 Sentence Reconstruction Performance

7 Additional Material

8 Summary and Conclusions

Privacy-Preserving Split Learning via Pareto Optimal Search

2.1 Adversarial Training-Based Defense

2.2 Information Theoretical Defense

3.1 Split Learning

3.2 Multi-task Learning and Multi-objective Optimization

4 Problem Formulation and Threat Model

5 Methodology

5.1 Privacy and Utility Losses

5.2 Trajectory-Controlled Pareto Optimal Search

5.3 Privacy-Preserving Training

6 Experiments and Evaluation

6.1 Experimental Setup

6.2 Effect of Trajectory-Controlled Pareto Optimal Search

6.3 Privacy Against Attacks

6.4 Privacy-Utility Tradeoffs

7 Conclusion

A Notations

B Proof for Loss Formulation

C Proof of Lemma 2

D Proof of Lemma 3

E Proof of Theorem 1

Security of NVMe Offloaded Data in Large-Scale Machine Learning

3 NVMevade - Untargeted Poisoning Attack on NVMe

3.1 System Model

3.2 Design of NVMevade

3.3 Evaluation

4 NVMevade - Preventing Poisoning Attacks on NVMe

4.1 Design of NVMensure

4.2 Instantiation of NVMevade for DeepSpeed

4.3 Evaluation

5 Related Work

A Instantiation for DeepSpeed

A.1 NVMevade

A.2 NVMensure

PassGPT: Password Modeling and (Guided) Generation with Large Language Models

1.1 Contributions and Remarks

2 Background and Related Work

2.1 Deep Generative Models

2.2 Transformers

2.3 Related Work

3 Experimental Setup

3.1 Datasets

3.2 Our Models

4.1 Password Generation

4.2 Generalizing to Longer Passwords and Unseen Distributions.

4.3 Guided Generation

4.4 Probabilities and Entropies Estimates by PassGPT

4.5 Discussion: PassGPT Vs PassVQT

4.6 Password Strength Estimation

5 Conclusions

A Passwords at the Quantiles of the Distribution

B PassGPT vs PassVQT

Learning Type Inference for Enhanced Dataflow Analysis

2.1 Code Property Graphs and Code Analysis

2.2 Statistical Type Inference

2.3 Large Language Models

3 Motivation

4 Design

4.1 Code Property Graph Usage Slicing

4.2 Machine Learning Model

4.3 Integration in Joern

5.1 Type Inference Generalization

5.2 Efficacy of JoernTI

6 Related Work

Efficient Pruning for Machine Learning Under Homomorphic Encryption

2 Related Work

3 Background

3.1 Homomorphic Encryption

3.2 Ciphertext Packing Using Tile Tensors

3.3 Threat Model

4 The HE-PEx Framework

4.1 Prune, Pack, Permute and Expand Methods

4.2 Layer Co-Permutation for Improved Pruning

4.3 Privacy Considerations

6 Evaluation

6.1 Analysis of Our Pruning Schemes

6.2 Comparison with P2T Strategies

6.3 Impact of Permutation

6.4 Impact on Latency and Memory

A.1 Additional Evaluation

A.2 Experimental Setup

Software and Systems Security

SoK: A Tale of Reduction, Security, and Correctness - Evaluating Program Debloating Paradigms and Their Compositions

2 Debloating Methods

2.1 Application-Level Program Debloating

2.2 Library-Level Program Debloating

3 Components of the DebloatBenchA Framework

3.1 Framework Components

4 Experimental Setup in DebloatBenchA

4.1 Standalone Mode

4.2 Composition Mode

5 Evaluation of Debloaters.

5.1 Evaluation Results on 10 Coreutil Programs

5.2 Evaluation Results on 5 Non-coreutils Programs.

5.3 RQ5: Debloater Composition

6 Discussion

6.1 Impact of Design Choices

6.2 Limitations of This Study

7 Related Work

8 Conclusion

On the (In)Security of Manufacturer-Provided Remote Attestation Frameworks in Android

2.1 Remote Attestation

2.2 Mobile Service Core

2.3 Integrity on Android

3 MARA Frameworks Demystification

3.1 MARA Scheme

3.2 Details About SafetyNet

3.3 Details About SafetyDetect

4 Bypassing MARA Protection

4.1 Scenarios and Attacker's Capabilities

4.2 Fundamental Observation

4.3 Bypassing Device Integrity Check

4.4 Bypassing App Integrity Check

4.5 Bypassing Data Integrity Protection

4.6 Attack Implementation

5.1 Effectiveness of Our Bypassing Approach

5.2 Large-Scale Measurement Study

A Integrity Checking Items

A.1 Device Integrity Checking Items

A.2 App Integrity Checking Items

B Bypassing app integrity check

C Device details

DScope: To Reliably and Securely Acquire Live Data from Kernel-Compromised ARM Devices

2.1 Background on ARM Virtualization

2.2 Related Work

3 The Design of DScope

3.1 Problem Description

3.2 The Design of DScope

3.3 The Hypervisor

3.4 Device Operation Environment (DOE)

3.5 Protection on DOE

4 Implementation

4.1 Hypervisor Implementation

4.2 Interrupt Handling

4.3 Device Operation Environment

4.4 DOE Agent

4.5 Case Studies

5.1 Reliability of DScope Launch

5.2 Reliability Experiments

5.3 Security and Reliability of DOE

5.4 Security Assessment

5.5 Performance Evaluation

References.

SplITS: Split Input-to-State Mapping for Effective Firmware Fuzzing.

Notes:

Includes bibliographical references and index.

Description based on publisher supplied metadata and other sources.

Description based on print version record.

Other Format:

Print version: Tsudik, Gene Computer Security - ESORICS 2023

ISBN:

9783031514821

OCLC:

1417668055