Exposing Digital Forgeries Through Chromatic Aberration Micah K. Johnson Hany Farid Department of Computer Science Department of Computer Science Dartmouth College Dartmouth College Hanover, NH 03755 Hanover, NH 03755 [email protected] [email protected] ABSTRACT cases, chromatic aberration leads to various forms of color Virtually all optical imaging systems introduce a variety of imperfections in the image. To a first-order approximation, aberrations into an image. Chromatic aberration, for exam- longitudinal aberration can be modeled as a convolution of ple, results from the failure of an optical system to perfectly the individual color channels with an appropriate low-pass focus light of different wavelengths. Lateral chromatic aber- filter. Lateral aberration, on the other hand, can be mod- ration manifests itself, to a first-order approximation, as an eled as an expansion/contraction of the color channels with expansion/contraction of color channels with respect to one respect to one another. When tampering with an image, another. When tampering with an image, this aberration is these aberrations are often disturbed and fail to be consis- often disturbed and fails to be consistent across the image. tent across the image. We describe a computational technique for automatically es- We describe a computational technique for automatically timating lateral chromatic aberration and show its efficacy estimating lateral chromatic aberration. Although we even- in detecting digital tampering. tually plan to incorporate longitudinal chromatic aberra- tion, only lateral chromatic aberration is considered here. We show the efficacy of this approach for detecting digital Categories and Subject Descriptors tampering in synthetic and real images. This work provides I.4 [Image Processing]: Miscellaneous another tool in a growing number of image forensic tools, [3, 7, 8, 10, 9, 5, 6]. General Terms 2. CHROMATIC ABERRATION Security We describe the cause of lateral chromatic aberration and derive an expression for modeling it. For purposes of expo- Keywords sition a one-dimensional imaging system is first considered, Digital Tampering, Digital Forensics and then the derivations are extended to two dimensions. 2.1 1-D Aberration 1. INTRODUCTION In classical optics, the refraction of light at the boundary Most images contain a variety of aberrations that result between two media is described by Snell’s Law: from imperfections and artifacts of the optical imaging sys- tem. In an ideal imaging system, light passes through the n sin(θ) = nf sin(θf ), (1) lens and is focused to a single point on the sensor. Optical where θ is the angle of incidence, θf is the angle of refrac- systems, however, deviate from such ideal models in that tion, and n and nf are the refractive indices of the media they fail to perfectly focus light of all wavelengths. The through which the light passes. The refractive index of glass, resulting effect is known as chromatic aberration which oc- nf , depends on the wavelength of the light that traverses it. curs in two forms: longitudinal and lateral. Longitudinal This dependency results in polychromatic light being split aberration manifests itself as differences in the focal planes according to wavelength as it exits the lens and strikes the for different wavelengths of light. Lateral aberration mani- sensor. Shown in Figure 1, for example, is a schematic show- fests itself as a spatial shift in the locations where light of ing the splitting of short wavelength (solid blue ray) and different wavelengths reach the sensor – this shift is pro- long wavelength (dashed red ray) light. The result of this portional to the distance from the optical center. In both splitting of light is termed lateral chromatic aberration. Lateral chromatic aberration can be quantified with a low-parameter model. Consider, for example, the position of short wavelength (solid blue ray) and long wavelength Permission to make digital or hard copies of all or part of this work for (dashed red ray) light on the sensor, xr and xb, Figure 1. personal or classroom use is granted without fee provided that copies are With no chromatic aberration, these positions would be not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to equal. In the presence of chromatic aberration, it can be republish, to post on servers or to redistribute to lists, requires prior specific shown (Appendix A) that these positions can be modeled permission and/or a fee. as: MM&Sec’06, September 26–27, 2006, Geneva, Switzerland. Copyright 2006 ACM 1-59593-493-6/06/0009 ...$5.00. xr ≈ αxb, (2) θ Lens θr f θb Sensor xr xb Figure 1: The refraction of light in one dimension. Figure 2: The refraction of light in two dimensions. Polychromatic light enters the lens at an angle θ, and Polychromatic light enters the lens and emerges at emerges at an angle which depends on wavelength. an angle which depends on wavelength. As a re- As a result, different wavelengths of light, two of sult, different wavelengths of light, two of which are which are represented as the red (dashed) and the represented as the red (dashed) and the blue (solid) blue (solid) rays, will be imaged at different points, rays, will be imaged at different points. The vec- xr and xb. tor field shows the amount of deviation across the image. where α is a scalar value. This model generalizes for any two wavelengths of light, where α is a function of these wave- lengths. is visible in high-contrast regions of an image as a magenta or green halo [4]. 2.2 2-D Aberration For a two-dimensional lens and sensor, the distortion 2.3 Estimating Chromatic Aberration caused by lateral chromatic aberration takes a form sim- ilar to Equation (2). Consider again the position of short In the previous section, a model for lateral chromatic aber- wavelength (solid blue ray) and long wavelength (dashed red ration was derived, Equations (4)–(5). This model describes the relative positions at which light of varying wavelength ray) light on the sensor, (xr, yr) and (xb, yb). In the pres- ence of chromatic aberration, it can be shown (Appendix A) strikes the sensor. With a three color channel RGB image, that these positions can be modeled as: we assume that the lateral chromatic aberration is constant within each color channel. Using the green channel as ref- (xr, yr) ≈ α(xb, yb), (3) erence, we would like to estimate the aberration between the red and green channels, and between the blue and green where α is a scalar value. channels. Deviations or inconsistencies in these models will Shown in Figure 2 is vector-based depiction of this aber- then be used as evidence of tampering. ration, where each vector ~v = `x − x , y − y ´. Note that r b r b Recall that the model for lateral chromatic aberration this model is simply an expansion/contraction about the consists of three parameters, two parameters for the cen- center of the image. In real lenses, the center of optical ter of the distortion and one parameter for the magnitude aberrations is often different from the image center due to of the distortion. These model parameters will be denoted the complexities of multi-lens systems [12]. The previous (x , y , α ) and (x , y , α ) for the red to green and blue to model can therefore be augmented with an additional two 1 1 1 2 2 2 green distortions, respectively. parameters, (x , y ), to describe the position of the expan- 0 0 The estimation of these model parameters can be framed sion/contraction center. The model now takes the form: as an image registration problem [1]. Specifically, lateral xr = α(xb − x0) + x0 (4) chromatic aberration results in an expansion or contraction y = α(y − y ) + y . (5) between the color channels, and hence a misalignment be- r b 0 0 tween the color channels. We, therefore, seek the model It is common for lens designers to try to minimize chro- parameters that bring the color channels back into align- matic aberration in lenses. This is usually done by combin- ment. There are several metrics that may be used to quan- ing lenses with different refractive indices to align the rays tify the alignment of the color channels. To help contend for different wavelengths of light. If two wavelengths are with the inherent intensity differences across the color chan- aligned, the lens is called an achromatic doublet or achro- nels we employ a metric based on mutual information that mat. It is not possible for all wavelengths that traverse an has proven successful in such situations [11]. We have found achromatic doublet to be aligned and the residual error is that this metric achieves slightly better results than a sim- known as the secondary spectrum. The secondary spectrum pler correlation coefficient metric (with no difference in the 700 600 500 400 300 Frequency 200 100 0 0 5 10 15 20 25 30 35 Average Angular Error (degrees) Figure 3: Synthetically generated images. Shown are, from left to right, a sample image, the distortion applied to the blue channel (the small circle denotes the distortion center), the estimated distortion, and a histogram of angular errors from 2000 images. For purposes of display, the vector fields are scaled by a factor of 50. run-time complexity1). Other metrics, however, may very displacement fields for these distortions are: well achieve similar or better results. „(α (x − x ) + x ) − x« We will describe the estimation of the red to green dis- ~v (x, y) = 0 0 0 (10) 0 (α (y − y ) + y ) − y tortion parameters (the blue to green estimation follows a 0 0 0 „ « similar form).