Inhalt

•  
   Zusammenfassung
•  
   Abstract
•  
   Acknowledgements
•  
   Contents
•  
   Introduction
•  
   Fast Convex Relaxations using Graph Discretizations
•  
   Introduction
•  
   Related Work
•  
  Graph Discretizations for Convex Relaxations
•  
   Preliminaries
•  
   Graph Discretization
•  
   Numerical Evaluation
•  
   Segmentation
•  
   Stereo Matching
•  
   Conclusions
•  
   Proof of Proposition 1
•  
   Algorithmic Details
•  
   Implementation
•  
   Experimental Setup
•  
   Superpixel-Sublabel Stereo Lifting
•  
   Further plots
•  
   Composite Optimization by Nonconvex Majorization-Minimization
•  
   Introduction
•  
   Related Work
•  
   Organization of this work
•  
  The General Principle
•  
   The Algorithm
•  
   Special Cases
•  
   Proximity relative to the inner function
•  
  Algorithm Discussion and Convergence
•  
   Basic Properties
•  
   Descent Properties
•  
   Convergence Properties
•  
   Global Convergence of the inner sequence
•  
  Implementation Details
•  
   Modeling
•  
   Choices for the Bregman Distance
•  
   An example of a non-separable, solvable subproblem
•  
   Inertia
•  
  Experimental results
•  
   Synthetic experiments
•  
   Time-of-Flight Depth Reconstruction
•  
   Conclusions
•  
   Reformulation as continuous Majorizer
•  
   Details regarding Grid Search
•  
   Parametric Majorization for Data-Driven Energy Minimization Methods
•  
   Introduction
•  
   Related Work
•  
  Bi-Level Learning
•  
   Majorization of Bi-level Problems
•  
   Single-Level Majorizers
•  
   Intermission: One-Dimensional Example
•  
   Iterative Majorizers
•  
  Examples
•  
   Computed Tomography
•  
   Variational Segmentation
•  
   Analysis Operator Models
•  
   Conclusions
•  
  Convex Analysis in Section 3
•  
   Details for Derivation of (4.11) to (4.12)
•  
   Details for Derivation of (4.14) to (4.15)
•  
   Proof of Proposition 2
•  
   Derivation of the surrogate functions for the example in 4.3.3
•  
   Proof of Proposition 4
•  
  Experimental Setup
•  
   CT - Additional Details
•  
   Segmentation - Additional Details
•  
   Analysis Operators - Additional Details
•  
   Extended Overview of Related Work
•  
   Analysis Operator Learning - Additional Figures
•  
   Derivation from Support Vector Machine Principles
•  
   On Hessian Inversion
•  
   Generalization of the Iterative Surrogate
•  
   Inverting Gradients - How easy is it to break privacy in federated learning?
•  
   Introduction
•  
   Related Work
•  
   Theoretical Analysis: Recovering Images from their Gradients
•  
   A Numerical Reconstruction Method
•  
   Single Image Reconstruction from a Single Gradient
•  
   Distributed Learning with Federated Averaging and Multiple Images
•  
   Conclusions
•  
   Broader Impact - Federated Learning does not guarantee privacy
•  
  Variations of the threat model
•  
   Dishonest Architectures
•  
   Dishonest Parameter Vectors
•  
   Experimental Details
•  
  Hyperparameter Settings
•  
   Settings for the experiments in Sec.5.5
•  
   Setting for experiments in Sec. 5.6
•  
   Proofs for section 5.2
•  
  Additional Examples
•  
   Additional CIFAR-10 examples
•  
   Visualization of experiments in Sec. 5.5
•  
   More ImageNet examples for Sec. 5.5
•  
   Multi-Image Recovery of Sec. 5.6
•  
   General case of Sec. 5.6
•  
   Witches' Brew: Industrial Scale Data Poisoning via Gradient Matching
•  
   Introduction
•  
   Related Work
•  
  Efficient Poison Brewing
•  
   Threat Model
•  
   Motivation
•  
   The Central Mechanism: Gradient Alignment
•  
   Making attacks that transfer and succeed ``in the wild''
•  
   Theoretical Analysis
•  
  Experimental Evaluation
•  
   Evaluations on CIFAR-10
•  
   Poisoning ImageNet models
•  
   Deficiencies of Defense Strategies
•  
   Conclusion
•  
   Remarks
•  
   Experimental Setup
•  
   Hardware
•  
   Models
•  
   Cloud AutoML Setup
•  
   Proof of Proposition 6.1
•  
   Poisoned Datasets
•  
   Visualizations
•  
  Additional Experiments
•  
   Full-scale MetaPoison Comparisons on CIFAR-10
•  
   Deficiencies of Filtering Defenses
•  
   Details: Defense by Differential Privacy
•  
   Details: Gradient Alignment Visualization
•  
   Ablation Studies - Reduced Brewing/Victim Training Data
•  
   Ablation Studies - Method
•  
   Transfer Experiments
•  
   Multi-Target Experiments
•  
   Conclusions
•  
   References
•  
   Index
•  
   Lists of Figures and Tables