Dynamic particle swarm optimization (DPSO) of recurrent problems

Dynamic particle swarm optimization (DPSO) of recurrent problems

GENERAL INTRODUCTION

 Numerous applications require the storage and transmission of document images. This leads to serious privacy concerns, specially considering the sensitive nature of the data stored in such images. Enforcing the security of document images is a paramount issue for many industries including financial, medical and legal. One easy strategy to enforce the security of document images is by means of cryptography. Howeve r, once an image has been decrypted, it can be easily manipulated and transmitted. Avoiding abuse (specially by insiders) requires a security mechanism that will “follow” the image wherever it goes and no matter what manipulation it suffers (as long as the manipulation does not affect its commercial value). Digital watermarking allows the embedding of image-related data in a covert manner by manipulation of pixel values. This process is subject to a trade-off between robustness against image processing operations (attacks) and image quality. Since it is covert and involves manipulation of pixel values, a watermark provides means of enforcing the integrity and authenticity of a given image. The common approach is to employ a robust watermark (which can resist attacks) in order to enforce authenticity and a fragile watermark (which is easily destroyed by attacks) in order to detect tampering (enforce integrity).

Evaluation of robustness

  As mentioned before, robustness refers to the ability to detect the watermark after the watermarked image has suffered common signal processing operations. These operations can be intentional or not. The intentional use of such type of operation in awatermarked image is called an attack. There are four main families of attacks: removal, geometric, cryptographic and protocol attacks (Voloshynovskiy et al., 2001). In a removal attack, the embedded watermark is partially or completely removed either by a source of noise or with the use of image processing techniques such as de-noising, lossy compression, cropping, etc. In a geometric attack by its way, the watermark is not removed but instead, the synchronization between the embedder and detector is affected with the use of affine transformations, such as rotation. In a cryptographic attack, the intention is to crack the security mechanisms employed on watermarking (such as the watermarking key). Finally, in a protocol attack, the objective is to threaten the validity of the system rather than its functionality. For example, in an protocol attack known as invertible watermark, an attacker extracts his own watermark from a watermarked image and claims he is the owner.

  Intelligent watermarking usually aims at improving the robustness against remova l attacks, since it is possible to increase the robustness against such attacks by adjusting embedding parameters (at the cost of adding more visible artifacts). Geometric attacks can be addressed either by detecting and inverting the distortion in the detector (Wu, 2001; Cox et al., 2002) or by embedding the data in a domain resistant to affine transformations such as the Discrete Fourier Transform (DFT) (ÓRuanaidh and Pun, 1998). Cryptographic attacks can be made unfeasible by using large watermark keys. Finally, protocol attacks can be minimized by embedding signal-dependent watermarks, for example, a signature of the cover work(Yang and Kot, Dec. 2006).

Challenges of watermarking

  The use of digital watermarks makes possible the embedding of side information into a cover image in an imperceptible way. The embedding must be performed according to a trade-off between robustness and image quality. Watermarking can thus, be considered an optimization problem. The main advantage of securing a document image with a digital watermark is that theprotection provided is not ostensive. Depending on the perceptual model employed, the authenticity and integrity of a  document are protected in an invisible manner. Despite these advantages, there are many known attacks to digital watermarking systems. For example, if a watermark  detector is widely available, an attacker could use detection information to repeatedly make small changes to the watermarked work until the detector fails to detect the watermark (Muharemagic, 2004). Moreover, in a type of attack named ambiguity attack, someone can add a watermark to an already watermarked work in such a way that it would appear that this second watermark is the true watermark. In another type of attack named geometric attack, rotation, scale and translation transformations are applied to the watermarked image in a way that the synchronization between the embedded and detected watermark signal is lost, what could be a threat for an authenticity application.

   The use of a robust watermark can mitigate the effects of most of these attacks (except for geometric attacks, which must be tackled with the use of registration marks (Cox et al., 2002;Wu, 2001)), at the expense of adding more visual artifacts. This makes robust watermarks very attractive for authenticity applications. A fragile watermark can be very useful in the detection  of intentional or unintentional modifications in the cover image (integrity enforcement). A watermark can be added in a fragile manner, and once its detection fails, it can be assumed that the image was tampered. The side effect of using fragile watermarks is that its detection will be affected by small variations in the image due to compression, processing or channel noise (here the cover image is considered a source of noise to the watermark signal). A balance between tampering protection and noise robustness must be considered.

Table des matières

GENERAL INTRODUCTION
CHAPTER 1 INTELLIGENT WATERMARKING
1.1 Introduction
1.2 Digital Watermarking
1.2.1 Survey of Watermarking Techniques
1.2.1.1 Embedding effectiveness
1.2.1.2 Fidelity
1.2.1.3 Embedding rate
1.2.1.4 Blind or informed embedding
1.2.1.5 Informed coding
1.2.1.6 Reliability
1.2.1.7 Robustness
1.2.1.8 Bi-tonal images
1.2.1.9 Embedder
1.2.1.10 Detector
1.2.2 Evaluation of visual impact
1.2.3 Evaluation of robustness
1.2.4 Challenges of watermarking.
1.3 Intelligent watermarking 
1.3.1 Supervised learning
1.3.1.1 MLP
1.3.1.2 SVM
1.3.2 Optimization of watermarking parameters
1.3.3 Key Issues
1.4 Case study – optimization of a bi-tonal watermarking system using PSO
1.4.1 Framework
1.4.1.1 Baseline watermarking system
1.4.1.2 Particle Swarm Optimization (PSO)
1.4.2 Simulation results
1.4.2.1 Baseline adaptive system (Muharemagic, 2004)
1.4.2.2 Adaptive system based on PSO
1.4.2.3 Discussion
1.5 Conclusion 
1.6 Discussion
CHAPTER 2 HIGH THROUGHPUT INTELLIGENT WATERMARKING OF HOMOGENEOUS STREAMS OF BI-TONAL IMAGES
2.1 Introduction .
2.2 Digital watermarking methods for bi-tonal images
2.3 Intelligent watermarking of isolated images using Particle Swarm Optimization (PSO) 
2.4 Fast intelligent watermarking of image streams using Dynamic PSO
2.4.1 Change detection
2.4.2 A memory-based intelligent watermarking method using DPSO
2.5 Experimental results 
2.5.1 Methodology
2.5.1.1 Database
2.5.2 A – Optimization of isolated bi-tonal images using full PSO versus default
embedding parameters
2.5.3 B – Optimization of streams of bi-tonal images using memory-based DPSO
versus full PSO
2.5.3.1 No attack
2.5.3.2 Attack modeling – cropping of 1% of image surface
2.5.4 C – Optimization of streams of bi-tonal images using memory-based DPSO
(learning mode) versus full PSO
2.5.5 Discussion
2.6 Conclusion 
2.7 Discussion
CHAPTER 3 FAST INTELLIGENT WATERMARKING OF HETEROGENEOUS IMAGE 
3.1 Introduction 
3.2 Optimization problem formulation of intelligent watermarking 
3.3 Related work
3.3.1 Dynamic particle swarm optimization (DPSO) of recurrent problems
3.3.2 Pattern classification
3.4 Fast intelligent watermarking using Gaussian modeling of PSO populations 
3.4.1 What to store?
3.4.2 How to organize and update? .
3.4.2.1 Memory management operators – insert, merge and delete
3.4.3 How to retrieve solutions?
3.5 Simulation results 
3.5.1 Experimental protocol
3.5.1.1 Databases
3.5.1.2 Methodology
3.5.2 Overview
3.5.3 Scenario A – optimization of heterogeneous streams of bi-tonal images using memory-based DPSO versus full PSO
3.5.3.1 LTM fill up
3.5.3.2 Adaptive memory management
3.5.3.3 Impact of choice of confidence level .
3.5.3.4 Memorization performance
3.5.3.5 Other attacks
3.5.3.6 Adaptation performance
3.5.4 Scenario B – optimization of homogeneous streams of bi-tonal images using
memory-based DPSO versus full PSO
3.5.5 Scenario C – optimization of unconstrained (homogeneous/heterogeneous)
streams of bi-tonal images using memory-based DPSO versus full PSO
3.5.6 Discussion
3.6 Conclusion 
3.7 Discussion
CHAPTER 4 DS-DPSO: A DUAL SURROGATE APPROACH FOR INTELLIGENT WATERMARKING OF BI-TONAL DOCUMENT IMAGE STREAMS
4.1 Introduction 
4.2 Particle swarm optimization of embedding parameters
4.3 Surrogate-based optimization
4.4 A dual-surrogate DPSO approach for fast intelligent watermarking
4.4.1 System overview
4.4.2 STM and LTM recall
4.4.3 Off-line/on-line surrogate PSO
4.4.3.1 On-line update of GMMs
4.4.3.2 Gaussian Mixture Regression (GMR)
4.4.3.3 Evolution control
4.4.3.4 Off-line surrogate PSO
4.4.3.5 On-line surrogate PSO
4.5 Experimental methodology 
4.6 Simulation results
4.6.1 Case I – adaptation performance
4.6.2 Case II – comparison to previous DPSO approach (Vellasques et al. , 2012b)
4.6.2.1 Heterogeneous streams
4.6.2.2 Homogeneous streams
4.6.3 Case III – memorization capacity
4.6.4 Case IV – management of different attacks
4.6.5 Discussion
4.7 Conclusion
GENERAL CONCLUSION
APPENDIX I BASELINE BI-TONAL WATERMARKING SYSTEM
APPENDIX IIEMPIRICAL RUNTIME PERFORMANCE
BIBLIOGRAPHY

Télécharger le rapport complet