Estimation of Copy-sensitive Codes Using a Neural Approach

Abstract

Copy sensitive graphical codes are used as anti-counterfeiting solution in packaging and document protection. Their security is funded on a design hard-to-predict after print and scan. In practice there exist different designs. Here random codes printed at the printer resolution are considered. We suggest an estimation of such codes by using neural networks, an in-trend approach which has however not been studied yet in the present context. In this paper, we test a state-of-the-art architecture efficient in the binarization of handwritten characters. The results show that such an approach can be successfully used by an attacker to provide a valid counterfeited code so fool an authentication system. CCS CONCEPTS • Security and privacy → Authentication; • Computing methodologies → Neural networks.

Copy Sensitive Graphical Code Estimation: Physical vs Numerical Resolution

“…Obviously, if the resolution of the scanner is equal to that of the printer, the printed-and-scanned units have the same size as the printed ones: u = v. An opponent can try to estimate and retrieve the original binary CSGC by scanning the genuine printed version before re-printing in order to produce a fake one. Such an estimation attack consists of the following steps: 1) scanning at a resolution v o , where v o ≥ u, 2) image processing including a binarization process (as digital printers available on the market can only print black-and-white images), using a statistical approach [5] or machine learning techniques as SVM, LDA [3] and neural networks [6], [7], 3) printing the estimated CSGC, I at the same print resolution as the generation stage. The considered authentication system is presented in Fig.…”

“…The underlying process basically requests processing each input pixel for classification. Taking advantage of recent advancements in deep networks, the authors of [7] suggested the use of a fully convolutional auto-encoder that has previously demonstrated promising results in binarization of degraded manuscripts [11]. Additional use of residual connection in the architecture provides an opportunity to efficiently train the deep network.…”

“…upscaling. For Selectional Auto-Encoder (SAE) [7] crop window was set to 128 × 128 pixels and size of kernel was set to 4 × 4 when v = 4, 8 × 8 when v = 8 and 16 × 16 when v = 16. For SRGAN we defined l V GG/11.10 with φ 11,10 , a loss defined on very deep feature maps.…”

“…1.b, it can severely impact the CSGC structure. Nevertheless some statistical learning techniques [3], [4], [5] have shown that CSGC can be efficiently estimated after P&S. Such vulnerabilities are even more salient when the learning stage makes use of some artificial neural networks [6], [7]. In this paper, we consider a generative model to reverse engineer the P&S process.…”

“…Either the dedicated layers are used (we will name the network, SRGANb) or not (we will name the network, srGANb) such that we can compare physical resolution and numerical resolution with globally the same architecture. At any of these resolutions, the results on srGANb are compared with those of the state-of-the-art, namely with Selectional Auto-Encoder (SAE) which performs well CSGC decoding [7]. Learning GAN architectures as SRGANb or srGANb, is learning a pair of networks which are in competition with each other to minimize the loss such that the generated data are as similar as possible to real data.…”