• Share
  • Email
  • Embed
  • Like
  • Save
  • Private Content
Digital Image Authentication From JPEG Headers
 

Digital Image Authentication From JPEG Headers

on

  • 753 views

services on...... embedded(ARM9,ARM11,LINUX,DEVICE DRIVERS,RTOS) VLSI-FPGA DIP/DSP PLC AND SCADA JAVA AND DOTNET iPHONE ANDROID If ur intrested in these project please feel free to contact ...

services on...... embedded(ARM9,ARM11,LINUX,DEVICE DRIVERS,RTOS) VLSI-FPGA DIP/DSP PLC AND SCADA JAVA AND DOTNET iPHONE ANDROID If ur intrested in these project please feel free to contact us@09640648777,Mallikarjun.V

Statistics

Views

Total Views
753
Views on SlideShare
753
Embed Views
0

Actions

Likes
0
Downloads
0
Comments
0

0 Embeds 0

No embeds

Accessibility

Categories

Upload Details

Uploaded via as Adobe PDF

Usage Rights

© All Rights Reserved

Report content

Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

Cancel
  • Full Name Full Name Comment goes here.
    Are you sure you want to
    Your message goes here
    Processing…
Post Comment
Edit your comment

    Digital Image Authentication From JPEG Headers Digital Image Authentication From JPEG Headers Document Transcript

    • 1066 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 6, NO. 3, SEPTEMBER 2011 Digital Image Authentication From JPEG Headers Eric Kee, Micah K. Johnson, and Hany Farid Abstract—It is often desirable to determine if an image has somewhat distinct, and can, therefore, be used to determine if anbeen modified in any way from its original recording. The JPEG image has been altered from its original recording. Building onformat affords engineers many implementation trade-offs which this earlier work, we describe how various aspects of the JPEGgive rise to widely varying JPEG headers. We exploit these vari-ations for image authentication. A camera signature is extracted format can be used for authentication. Unlike previous work,from a JPEG image consisting of information about quantization this approach considers several features of the JPEG formattables, Huffman codes, thumbnails, and exchangeable image file not previously considered, namely properties of the run-lengthformat (EXIF). We show that this signature is highly distinct encoding employed by the JPEG standard, and aspects ofacross 1.3 million images spanning 773 different cameras andcell phones. Specifically, 62% of images have a signature that is the EXIF header format. This analysis is validated on overunique to a single camera, 80% of images have a signature that 1.3 million images spanning 33 different camera manufacturersis shared by three or fewer cameras, and 99% of images have a and 773 different camera and cell phone models.signature that is unique to a single manufacturer. The signatureof Adobe Photoshop is also shown to be unique relative to all II. METHODS773 cameras. These signatures are simple to extract and offer anefficient method to establish the authenticity of a digital image. The JPEG file format has emerged as a universal image stan- dard employed by nearly all commercial digital cameras [30], Index Terms—Digital forensics, digital tampering. [31]. As such, we consider the details of this encoding scheme, and how these details vary among cameras of different make, model, resolution, and quality. Cameras often support multiple I. INTRODUCTION resolutions and qualities, each of which yield images with dif-D IGITAL images are now routinely introduced as evidence ferent JPEG compression parameters. We will, therefore, ex- into courts of law. It has, therefore, become critical to plore the JPEG parameters used by cameras of different makeverify the integrity of this digital evidence. Digital forensic tech- and model, and by each camera under different resolution andniques have been developed to detect various traces of tam- quality settings.pering: region duplication [1]–[4]; resampling [5], [6]; color We begin by describing the JPEG compression standard andfilter array artifacts [7], [8]; inconsistencies in camera response file format. Given a three-channel color image (RGB), compres-function [9]; inconsistencies in lighting and shadows [10]–[13]; sion proceeds as follows. An image is first transformed frominconsistencies in chromatic aberrations [14], [15]; inconsisten- RGB into luminance/chrominance space (YCbCr). The twocies in sensor noise [16], [17]; and inconsistencies in statistical chrominance channels (CbCr) are typically subsampled by afeatures [18]. See [19] for a general survey. However, relatively factor of two relative to the luminance channel (Y). Each channelbenign modifications either cannot be detected by these tech- is then partitioned into 8 8 pixel blocks. These values areniques, or render these techniques ineffective. converted from unsigned to signed integers (e.g., from [0,255] to It is often desirable to determine if a digital image has been [ 128,127]). Each block is converted to frequency spacealtered in any way from the time of its recording, including using a 2-D discrete cosine transform (DCT)manipulations as simple as cropping. Previous work in de-tecting double JPEG compression holds some promise to detectany form of image manipulation [20]–[24]. These techniquesrequire fairly sophisticated models and analysis schemes, can (1)be vulnerable to simple countermeasures such as additive noiseor down-sampling, and can be computationally intensive. In where denotes a specific image channel, is a normalizingcontrast, it has been previously shown that exchangeable image scale factor, and is the underlying pixel values. Dependingfile format (EXIF) headers [25] and JPEG quantization tables on the specific frequency and channel , each DCT coeffi-[26]–[29] used by cameras and software manufacturers are cient is quantized by an amount Manuscript received November 16, 2010; revised March 08, 2011; accepted (2)March 08, 2011. Date of publication March 17, 2011; date of current versionAugust 17, 2011. This work was supported by a gift from Adobe Systems, Inc. This stage is the primary source of data reduction and informa-and a gift from Microsoft, Inc. The associate editor coordinating the review of tion loss.this manuscript and approving it for publication was Prof. Nasir Memon. E. Kee and H. Farid are with the Department of Computer Science, Dartmouth With some variations, the above sequence of steps is em-College, Sudikoff Laboratory, Hanover, NH 03755 USA (e-mail: erickee@cs. ployed by all JPEG encoders. The primary source of variationdartmouth.edu; farid@cs.dartmouth.edu). in these encoders is the choice of quantization values . The M. K. Johnson is with the Computer Science and Artificial Intelligence Lab-oratory, Massachusetts Institute of Technology, Cambridge, MA 02139 USA quantization is specified as a set of three 8 8 tables associ-(e-mail: kimo@csail.mit.edu). ated with each frequency and image channel (YCbCr). For low Digital Object Identifier 10.1109/TIFS.2011.2128309 compression rates, the values in these tables tend towards 1, and 1556-6013/$26.00 © 2011 IEEE
    • KEE et al.: DIGITAL IMAGE AUTHENTICATION FROM JPEG HEADERS 1067Fig. 1. Camera signatures for a Canon EOS Rebel XTi and Nikon D40 highlighting the differences (image dimensions, image quantization tables, thumbnailquantization tables, and EXIF counts) and similarities (thumbnail dimensions and image and thumbnail Huffman codes).increase for higher compression rates. Shown in the top por- are the tables for the luminance and chrominance channels. As istion of Fig. 1, for example, are the quantization tables employed typical, the quantization for the luminance channel is less than forby a Canon and Nikon digital camera. Shown, from top to bottom, the two chrominance channels, and the quantization is less for the
    • 1068 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 6, NO. 3, SEPTEMBER 2011lower frequency components (the frequency in each table is spec- the full-resolution image. The thumbnail is then typically com-ified in standard zig-zag order with the lowest frequency in the top pressed and stored in the header as a JPEG image. As such, weleft corner, and the highest frequency in the bottom right corner). can extract the same components from the thumbnail as from the After quantization, the DCT coefficients are subjected to en- full resolution image described in the previous section. As withtropy encoding (typically Huffman coding). Huffman coding is the full resolution image, we have found that the chrominancea variable-length encoding scheme that encodes frequently oc- channels are encoded with the same parameters.curring values with shorter codes, and less frequently occurring Some camera manufacturers do not create a thumbnail image,values with longer codes. This lossless compression scheme ex- or do not encode them as a JPEG image. In such cases, weploits the fact that the quantization of DCT coefficients yields simply assign a value of zero to all of the thumbnail parame-many zero coefficients, which can in turn be efficiently encoded. ters. Rather than being a limitation, we consider the lack of aMotivated by the fact that the statistics of the dc and ac DCT co- thumbnail as a characteristic property of a camera.efficients are different,1 the JPEG standard allows for different In total, we extract 284 values from the thumbnail image:Huffman codes for the dc and ac coefficients. This entropy en- 2 thumbnail dimensions, 192 quantization values, and 90coding is applied separately to each YCbCr channel, employing Huffman codes.separate codes for each channel. The JPEG standard does not enforce any specific quantiza- C. EXIF Metadata Parameterstion table or Huffman code. Camera and software engineers are, The final component of our camera signature is extractedtherefore, free to balance compression and quality to their own from an image’s EXIF metadata [32]. The metadata, found in theneeds and tastes. The specific quantization tables and Huffman JPEG header, stores a variety of information about the cameracodes needed to decode a JPEG file are embedded into the JPEG and image. According to the EXIF standard, there are five mainheader. In the following sections, we describe how the JPEG image file directories (IFDs) into which the metadata is orga-quantization table and Huffman codes along with other data ex- nized: 1) Primary; 2) Exif; 3) Interoperability; 4) Thumbnail;tracted from the JPEG header form a distinct camera signature and 5) GPS. Camera manufacturers are free to embed any (orwhich can be used for authentication. no) information into each IFD. We extract a compact represen- tation of their choice by counting the number of entries in eachA. Image Parameters of these five IFDs. The first three components of our camera signature are the Because the EXIF standard allows for the creation of addi-image dimensions, quantization table, and Huffman code. The tional IFDs, we also count the total number of any additionalimage dimensions are used to distinguish between cameras with IFDs, and the total number of entries in each of these. Somedifferent sensor resolution. In order to compensate for landscape camera manufacturers customize their metadata in ways that doand portrait images, the image dimensions are specified as the not conform to the EXIF standard. These customizations yieldminimum dimension followed by the maximum dimension. The errors when parsing the metadata. We consider these errors to beset of three 8 8 quantization tables are specified as a one-di- a feature of camera design and, therefore, count the total numbermensional array of 192 values: each channel’s table is specified of parser errors, as specified by the EXIF standard.in column-order, and the three tables are specified in the order In total, we extract 8 values from the metadata: 5 entry countsof luminance (Y), chrominance (Cb), and chrominance (Cr). from the standard IFDs, 1 for the number of additional IFDs, 1 The Huffman code is specified as six sets of 15 values cor- for the number of entries in these additional IFDs, and 1 for theresponding to the number of codes of length 1, 2, , 15: each number of parser errors.of three channels requires two codes, one for the dc coefficients D. Image Authenticationand one for the ac coefficients. This representation eschews theactual code for a more compact representation that distinguishes As described in the previous sections, we extract 284 headercodes based on the distribution of code lengths. values from the full resolution image, a similar 284 header In theory the chrominance channels, Cb and Cr, can employ values from the thumbnail image, and another 8 from the EXIFdifferent quantization values and Huffman codes. In all of the metadata, for a total of 576 values. These 576 values form thecameras analyzed, however, we have found that the chromi- signature by which images will be authenticated. Specifically,nance channels are encoded with the same parameters. the signature and camera make and model are extracted from In total, we extract 284 values from the full resolution image: the EXIF metadata and compared against authentic image2 image dimensions, 192 quantization values, and 90 Huffman signatures extracted from the same camera make and model.codes. To the extent that photo-editing software will employ JPEG parameters that are distinct from the camera’s, any manipulationB. Thumbnail Parameters will alter the original signature, and can, therefore, be detected. Specifically, photo alteration is detected by extracting the sig- A thumbnail version of the full resolution image is often nature from an image and comparing it to a database of knownembedded in the JPEG header. The next three components of authentic camera signatures. Any matching camera make andour camera signature are extracted from this thumbnail image. A model can then be compared to the make and model specifiedthumbnail is typically no larger in size than a few hundred square in the image’s EXIF metadata. Any mismatch is strong evidencepixels, and is created by cropping, filtering, and down-sampling of some form of tampering. 1The dc term refers to the (0,0) frequency in the top left corner of each quan- For ballistic purposes, an extracted signature can be comparedtization table. The ac terms refer to the remaining frequencies. against authentic signatures to determine which, if any, cameras
    • KEE et al.: DIGITAL IMAGE AUTHENTICATION FROM JPEG HEADERS 1069have matching signatures. This application may seem unneces-sary since the camera make and model are specified in the EXIFmetadata. However, an image’s EXIF metadata can be relativelyeasily edited to alter the camera make and model, so the signatureis a more reliable determinant of a camera’s source. Given that an image’s EXIF metadata can be easily edited, itmay seem peculiar to rely on it for forensic and ballistic pur-poses. It should be noted that modifying the content of any ex- Fig. 2. Shown are, in sorted order, the number of images collected for each ofisting EXIF field will not affect the extracted EXIF counts, and the 9163 camera configurations (on a logarithmic horizontal axis).hence will not affect the extracted signature. If, however, thecamera make and model fields are changed in an attempt to con-ceal the camera source, then this can be detected when the ex- 6) Under certain conditions, Flickr resizes images to a max-tracted signature is found to be inconsistent with the make and imum dimension of 640, 800, 1024, 1280, 1600, or 2048.model. Additionally, if the metadata is accidentally or intention- When doing so, we found that Flickr frequently employsally deleted, then the remaining portion of the signature can still one of two image quantization tables (Appendix I). Imagesbe used for authentication. If, on the other hand, the number were eliminated if they were of any of these sizes and em-of EXIF fields are increased or decreased, then the EXIF count ployed either of these quantization tables.will be inconsistent with the expected signature. Although an 7) Images whose resolution is not native to the camera makeimage’s EXIF metadata can be edited, it still provides useful in- and model were eliminated. The native resolutions of 1578formation for forensic and ballistic analysis. different cameras were gathered from the website dpre- The usefulness of the camera signature is only as good as our view.com using Amazon’s crowd-sourcing tool Mechan-ability to extract signatures from a multitude of cameras of dif- ical Turk. Users on Mechanical Turk were given a cameraferent make, model, quality, and resolution settings. Therefore, make and model and a link to dpreview.com. They werewe describe next the construction of a large database of images asked to retrieve the camera’s native resolutions from thecollected from online sources. manufacturer’s announcement for the camera or from the camera specification page. The data for each camera wasE. Image Database considered valid after three independent users entered the same list of image sizes. Approximately 10 million images were downloaded from the These filters reduced the original 10 million images to 2.2 mil-popular photo-sharing website Flickr.com. Since we are inter- lion images.ested in extracting original camera signatures, it was neces- The camera make, model, and signature were then extractedsary to eliminate any images that had been edited or altered by from each of these image’s metadata. In order to further elim-photo-editing software. To begin, only images tagged as “orig- inate possible edited or altered images, only those entries withinal” by Flickr were downloaded. The images were then sub- 25 or more images having the same paired make, model, andjected to a series of filtering stages: signature were retained. This yielded approximately 1.3 million 1) Images that were not three-channel color JPEG images images. These images span 9163 different distinct pairings of were eliminated. camera make, model, and signature and represent 33 different 2) Duplicate images, determined by comparing MD5 hashes, camera manufacturers and 773 different camera and cell phone were eliminated. models. We refer to a pairing of camera make, model, and sig- 3) Images with no metadata or reduced metadata were elimi- nature as a camera configuration. Because a camera can record nated. in multiple resolutions and qualities, it is often the case that the 4) Images with inconsistent metadata “modification” and same make and model camera can yield many different signa- “original” dates were eliminated tures. In our case, there is an average of 12 different signatures 5) Images with a metadata “software” tag, introduced or mod- for each camera make and model (the ratio of configurations to ified by a photo-editing software, were eliminated. The unique camera models is ). most common modifications were found to contain one of It is from these 1.3 million images and 9163 camera config- the following keywords: urations that the distinctness of the camera signature was ana- lyzed. Note that any search of this data will be computationally efficient as it will only need to be performed on the 9163 con- figurations, and not the much larger 1.3 million images. Finally, we note that many manufacturers employ different model names for the same camera. For example, the Canon Dig- ital IXUS is sold under the name Canon Powershot in the United States and Canada, and IXY Digital in Japan. These synonymous names are taken into consideration when collating the images. III. RESULTS Shown in Fig. 2 are the number of images, in sorted order, collected for each of 9163 camera configurations. The max-
    • 1070 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 6, NO. 3, SEPTEMBER 2011Fig. 3. Shown are the distributions of equivalence class sizes based on the distinctiveness of the signature (an equivalence class of size means that camerasof different make and model share the same signature). Panels (a) through (e) correspond to, from top to bottom, signatures consisting of image parameters only,thumbnail parameters only, EXIF parameters only, image and thumbnail parameters, and image, thumbnail, and EXIF parameters. The dashed vertical line in eachpanel corresponds to the maximum equivalence class size. The horizontal axis for panel (b) was truncated (the maximum equivalence class size is 960). See alsoFig. 4.imum image count (30 760) is for the Canon EOS Digital per camera configuration is 147.1 and 48, respectively. TheRebel XTi (also known as Canon EOS 400D Digital or EOS remaining configurations span cameras and cell phones fromKiss Digital X), and the mean and median number of images Apple, Asahi, Canon, Casio, Fuji, Hewlett-Packard, HTC, JVC,
    • KEE et al.: DIGITAL IMAGE AUTHENTICATION FROM JPEG HEADERS 1071 largest equivalence class is of size 188, with 2.8% of the camera configurations. Shown in Fig. 3(d) is the distribution of equivalence class sizes for the 568-valued signature based on the full resolution image and thumbnail image parameters: 24.9% of the camera configurations are in an equivalence class of size one, 15.3% are in an equivalence class of size two, 11.3% are in an equivalenceFig. 4. Shown are the percentage of camera configurations with an equivalence class of size three, 51.5% are in an equivalence class of sizeclass size of 1 5, and the median and maximum equivalence class size. Each three or less, and the largest equivalence class is of size 91, withrow corresponds to different subsets of the complete signature. See also Fig. 3. 1.3% of the camera configurations. Shown in Fig. 3(e) is the distribution of equivalence class sizes for the complete 576-valued signature: 69.1% of theKodak, Konica, Leica, LG, Minolta, Motorola, Nikon, Nokia, camera configurations are in an equivalence class of size one,Olympus, Panasonic, Pentax, Ricoh, Research In Motion 12.8% are in an equivalence class of size two, 5.7% are in an(Blackberry), Samsung, Sanyo, Sigma, Sony, Toshiba, Vivitar, equivalence class of size three, 87.6% are in an equivalenceand more. class of size three or less, and the largest equivalence class is Shown in Fig. 1 are the complete signatures for a Canon EOS of size 14, with 0.2% of the camera configurations. BecauseRebel XTi and a Nikon D40. From top to bottom are the image the distribution of cameras is not uniform, it is also usefuldimensions, image quantization tables, image Huffman codes, to consider the likelihood of an image, as opposed to camerathumbnail dimensions, thumbnail quantization tables, thumb- configuration, being in an equivalence class of size . Withnail Huffman codes, and EXIF counts. respect to the complete signature, 62.4% of images are in an The signatures from each of the 9163 camera configurations equivalence class of size one (i.e., are unique), 10.5% of imageswere compared to determine their distinctiveness. Specifically, are in an equivalence class of size two, 7.5% of images are inwe determine the distinctiveness across cameras of different an equivalence class of size three, and 80.4% of images are inmake and model and relative to photo-editing software. To an equivalence class of size three or less.begin, all cameras with the same signature were placed into Individually, the image, thumbnail and EXIF parameters arean equivalence class. An equivalence of size means that not particularly distinct [Fig. 3(a)–(c)], but when combined,cameras of different make and model share the same signature. they provide a highly distinct signature [Fig. 3(e)]. This sug-An equivalence class of size greater than 1 means that there is gests that the choice of parameters are not highly correlated,an ambiguity in identifying the camera make and model. We and hence their combination improves overall distinctiveness.would like to maximize the number of camera configurations Although the thumbnail parameters are the least distinct, theirin an equivalence class of size 1 and minimize the largest addition to the overall signature is significant as can be seen byequivalence class size. comparing Fig. 3(a) and (d), and the second and fourth rows of Shown in Fig. 3 and summarized in Fig. 4 are the distribution Fig. 4.of equivalence class sizes for the signature in its whole and in Shown in Fig. 5(a) are the relative contributions of the indi-its parts. vidual parameters. The horizontal bars denote the number of Shown in Fig. 3(a) is the distribution of equivalence class unique values for each feature. The EXIF count is most unique,sizes for the 284-valued signature based only on the full res- followed by image dimensions, image quantization table (Yolution image parameters: 12.9% of the camera configurations channel) and then thumbnail quantization table (Y channel).are in an equivalence class of size one (i.e., are unique), 7.9% The least distinct parameters are the Huffman tables and theare in an equivalence class of size two, 6.2% are in an equiva- thumbnail dimensions. Shown in Fig. 5(b)–(e) is this samelence class of size three, 27.0% are in an equivalence class of analysis for all Canon, Sony, Nikon, and Olympus cameras:size three or less, and the largest equivalence class is of size 185, the Canon cameras are considerably more consistent than otherwith 2.7% of the cameras configurations. manufacturers. This consistency will yield to ambiguities in Shown in Fig. 3(b) is the distribution of equivalence class identifying a camera based on its signature. As shown next,sizes for the 284-valued signature based only on the thumbnail these ambiguities are almost always manufacturer-specific.image parameters: 1.1% of the camera configurations are in an An equivalence class of size greater than 1 implies a nondis-equivalence class of size one, 1.1% are in an equivalence class tinct signature, and hence an ambiguity in identifying theof size two, 1.0% are in an equivalence class of size three, 3.2% camera make and model. We next consider the scope of thisare in an equivalence class of size three or less, and the largest ambiguity. Shown in Fig. 6 are the cameras’ make and modelequivalence class is of size 960, with 14.1% of the camera con- that are in the same equivalence class for the full 576-valuedfigurations. signature. Shown in the last two rows, for example, are the Shown in Fig. 3(c) is the distribution of equivalence class cameras in the largest equivalence classes of size of 11 and 14.sizes for the eight-valued signature based only on the EXIF Note that all of these cameras are variants of the Sony DSCmetadata parameters: 8.8% of the camera configurations are in series. Each of the four equivalence classes of size 10 consistan equivalence class of size one, 5.4% are in an equivalence exclusively of the Sony DSC series. Note that, because of theclass of size two, 4.2% are in an equivalence class of size three, multiple resolution and quality settings, the same camera can18.4% are in an equivalence class of size three or less, and the appear in multiple equivalence classes. Similarly, each of the
    • 1072 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 6, NO. 3, SEPTEMBER 2011Fig. 5. Shown in panel (a) are the relative number of unique signature elements for all 9163 camera configurations. The EXIF count is the most distinct and theHuffman codes are the least distinct. Shown below are the same statistics for the (b) Canon, (c) Sony, (d) Nikon, and (e) Olympus cameras. The Canon camerasare considerably more consistent than the other manufacturers. Each panel is normalized by the total number of camera configurations in the respective categories.seven equivalence classes of size 8 and each of the eleven In Motion (RIM) BlackBerry. Although not shown in Fig. 6,equivalence classes of size 7 consist exclusively of either the this pattern continues for nearly all equivalence classes of sizeSony DSC series, the Canon Powershot series, or the Research greater than 1. With only a few exceptions, each equivalence
    • KEE et al.: DIGITAL IMAGE AUTHENTICATION FROM JPEG HEADERS 1073Fig. 6. Each entry corresponds to an equivalence class of identical image, thumbnail, and EXIF parameters. The first column is the size of the equivalence class , the second column is the number of equivalence classes of size , and the third and fourth columns are the camera make and model. The Canon Powershotmodel is abbreviated as PS, and the Research In Motion BlackBerry is abbreviated RIM BB.class consists of cameras from the same manufacturer and camera, modifying the image, and then resaving the image withseries. The only exceptions are the following four equivalence the appropriate EXIF format and all of the appropriate parame-classes of size two: ters: image size, image quantization table, image Huffman code, 1) Casio EX-Z60 | Canon Powershot SX120 IS; thumbnail size, thumbnail quantization table, and thumbnail 2) Nikon Coolpix P90 | Panasonic DMC-FZ18; Huffman code. While this is certainly possible, it is currently 3) Panasonic DMC-TZ5 | Nikon Coolpix S52; beyond the scope of popular photo-editing software. Our anal- 4) Panasonic DMC-ZS7 | Nikon Coolpix S630. ysis is also vulnerable to a standard rebroadcast attack in which In summary, although there is an ambiguity in some of the a digital image is manipulated, printed, and rephotographed.signatures, the signature still significantly constrains the identity Because an image’s EXIF metadata can be easily edited,of the camera make and model. this analysis can provide a more reliable method to determine Lastly, the signature from Adobe Photoshop (versions 3, 4, camera make and model. This information can be useful in7, CS, CS2, CS3, CS4, CS5 at all qualities) were compared to other forensic analyses. For example, techniques for devicethe 9163 camera signatures. In this case, only the image and authentication (e.g., [16]) compare an image of unknownthumbnail quantization tables and Huffman codes were used for origin to a database of known cameras. For large databases,comparison. No overlap was found between any Photoshop ver- this analysis can be computationally demanding [33]–[35].sion/quality and camera manufacturer. As such, the Photoshop Reliable determination of camera make and model can reducesignatures, each residing in an equivalence class of size 1, are this complexity by focusing only on the relevant camera(s) inunique. This means that any photo-editing with Photoshop can the database.be easily and unambiguously detected. It is important to note that significant care must be taken when using images downloaded from photo-sharing websites such as Flickr. As described in Section II-E, we went to great lengths IV. DISCUSSION to ensure that the images being analyzed were not modified by We have shown that cameras produce distinct JPEG headers a user or by Flickr. Even with these precautions, it is possiblethat facilitate both forensic and ballistic analysis. This analysis that some edited or altered images slipped through our systemdoes not differentiate between benign and nefarious modifica- of checks. Anecdotally, we have noticed that JPEG images thattions. While this is a stringent criteria, it is useful in certain are initially captured in RAW format and then converted, in soft-arenas. This forensic analysis can be useful in a legal setting, ware, to JPEG can be difficult to distinguish from images thatfor example, where it is important to determine if evidence has are captured directly in JPEG format. Fortunately, it appears thatbeen altered in any way. many of these images can be filtered because RAW-to-JPEG As with any forensic technique, it is important to consider converters often create images with resolutions that are dif-countermeasures. In our case, a determined forger could con- ferent than the native camera resolutions. A more robust systemceal their traces of tampering by extracting the signature of a would require determining which JPEG quantization tables and
    • 1074 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 6, NO. 3, SEPTEMBER 2011Huffman codes are used by popular RAW conversion software, [9] Z. Lin, R. Wang, X. Tang, and H.-V. Shum, “Detecting doctored imagesand then eliminating any images that employ them. using camera response normality and consistency,” in Proc. Computer Vision and Pattern Recognition, San Diego, CA, 2005. We have focused on the JPEG standard because it remains [10] M. Johnson and H. Farid, “Exposing digital forgeries in complexthe most ubiquitous image format employed by digital cameras. lighting environments,” IEEE Trans. Inf. Forensics Security, vol. 3, no. 2, pp. 450–461, Jun. 2007.We note that should JPEG 2000 become more widely used, our [11] W. Zhang, X. Cao, J. Zhang, J. Zhu, and P. Wang, “Detecting photo-basic approach could be extended to this format. As with the graphic composites using shadows,” in Proc. IEEE Int. Conf. Multi-JPEG standard, the JPEG 2000 standard contains the same basic media and Expo, 2009, pp. 1042–1045. [12] C. Riess and E. Angelopoulou, “Scene illumination as an indicatorparameters that we have analyzed: an EXIF header, thumbnail, of image manipulation,” in Proc. 12th Int. Workshop on InformationHuffman coding, and quantization (of the wavelet, as opposed Hiding, Calgary, AB, Canada, 2010.to DCT, coefficients). [13] E. Kee and H. Farid, “Exposing digital forgeries from 3-D lighting en- vironments,” in Proc. Workshop on Information Forensics and Secu- The power of our forensic analysis lies in the ability to acquire rity, Seattle, WA, 2010.signatures from a wide variety of cameras and cell phones. This [14] M. Johnson and H. Farid, “Exposing digital forgeries through chro-poses significant challenges as new cameras and cell phones are matic aberration,” in Proc. ACM Multimedia and Security Workshop, Geneva, Switzerland, 2006.constantly released. We expect to continue building our data- [15] T. Gloe, A. Winkler, and K. Borowka, “Efficient estimation and large-base of images and camera information in order to keep up with scale evaluation of lateral chromatic aberration for digital image foren- sics,” in Proc. SPIE Conf. Media Forensics and Security, San Jose, CA,these continual changes. 2010. [16] J. Lukáš, J. Fridrich, and M. Goljan, “Digital camera identification from sensor noise,” IEEE Trans. Inf. Forensics Security, vol. 1, no. 2, pp. APPENDIX I 205–214, Jun. 2006. [17] M. Chen, J. Fridrich, M. Goljan, and J. Lukas, “Determining image Two quantization tables employed by Flickr. Shown from left origin and integrity using sensor noise,” IEEE Trans. Inf. Forensics Security, vol. 3, no. 1, pp. 74–90, Mar. 2008.to right are 8 8 quantization tables corresponding to the lumi- [18] M. Stamm and K. J. R. Liu, “Forensic detection of image manipulationnance (Y), and chrominance (Cb and Cr) channels. The values using statistical intrinsic fingerprints,” IEEE Trans. Inf. Forensics Se-correspond to the quantization applied to DCT coefficients, (2). curity, vol. 5, no. 3, pp. 492–506, Sep. 2010. [19] H. Farid, “A survey of image forgery detection,” IEEE Signal Process. Mag., vol. 26, no. 2, pp. 16–25, Mar. 2009. [20] A. Popescu and H. Farid, “Statistical tools for digital forensics,” in Proc. 6th Int. Workshop on Information Hiding, Toronto, Canada, 2004. [21] J. Zhang, H. Wang, and Y. Su, “Detection of double-compression in JPEG2000 images,” in Proc. Int. Symp. Intelligent Information Tech- nology Application, 2008, vol. 1, pp. 418–421. [22] Z. Qu, W. Luo, and J. Huang, “A convolutive mixing model for shifted double JPEG compression with application to passive image authen- tication,” in Proc. IEEE Int. Conf. Acoustics, Speech, and Signal Pro- cessing, 2008, pp. 1661–1664. [23] M. Sorell, “Conditions for effective detection and identification of pri- mary quantization of re-quantized JPEG images,” in Proc. Int. Conf. Forensic Applications and Techniques in Telecommunications, Infor- mation, and Multimedia, Adelaide, Australia, 2008. [24] B. Mahdian and S. Saic, “Detecting double compressed JPEG images,” in Proc. 3rd Int. Conf. Imaging for Crime Detection and Prevention, London, U.K., 2009. [25] P. Alvarez, “Using extended file information (EXIF) file headers in digital evidence analysis,” Int. J. Digital Evidence, vol. 2, no. 3, pp. 1–5, 2004. [26] H. Farid, Digital Image Ballistics From JPEG Quantization Department REFERENCES of Computer Science, Dartmouth College, Tech. Rep. TR2006-583, 2006. [1] J. Fridrich, D. Soukal, and J. Lukáš, “Detection of copy move forgery in [27] J. Kornblum, “Using JPEG quantization tables to identify imagery pro- digital images,” in Proc. Digital Forensic Research Workshop, Cleve- cessed by software,” Digital Investigation, vol. 5, pp. S21–S25, 2008. land, OH, Aug. 2003. [28] H. Farid, Digital Image Ballistics From JPEG Quantization: A Fol- [2] A. Popescu and H. Farid, Exposing Digital Forgeries by Detecting Du- lowup Study Department of Computer Science, Dartmouth College, plicated Image Regions Department of Computer Science, Dartmouth Tech. Rep. TR2008-638, 2008. College, Tech. Rep. TR2004-515, 2004. [29] E. Kee and H. Farid, “Digital image authentication from thumbnails,” [3] S. Bayram, H. T. Sencar, and N. Memon, “A survey of copy-move in Proc. SPIE Symp. Electronic Imaging, San Jose, CA, 2010. forgery detection techniques,” in IEEE Western New York Image Pro- [30] Digital Compression and Coding of Continuous-Tone Still Images Part cessing Workshop, Rochester, NY, 2008. 1: Requirements and Guidelines, ISO/IEC JTC1 Draft International [4] X. Pan and S. Lyu, “Region duplication detection using image fea- Standard 10918-1, 1991. ture matching,” IEEE Trans. Inf. Forensics Security, vol. 5, no. 4, pp. [31] G. Wallace, “The JPEG still picture compression standard,” Commun. 857–867, Dec. 2010. ACM, vol. 34, no. 4, pp. 30–44, 1991. [5] A. Popescu and H. Farid, “Exposing digital forgeries by detecting [32] Exchangeable Image File Format for Digital Still Camera: Exif Version traces of resampling,” IEEE Trans. Signal Process., vol. 53, no. 2, pt. 2.2. Japan Electronics and Information Technology Industries Associ- 2, pp. 758–767, Feb. 2005. ation, 2002. [6] M. Kirchner and T. Gloe, “On resampling detection in re-compressed [33] M. Goljan, J. Fridrich, and T. Filler, “Large scale test of sensor finger- images,” in Proc. IEEE Workshop on Information Forensics and Secu- print camera identification,” in Proc. SPIE Conf. Media Forensics and rity, 2009, pp. 21–25. Security, San Jose, CA, 2009. [7] A. C. Popescu and H. Farid, “Exposing digital forgeries in color filter [34] M. Goljan, J. Fridrich, and T. Filler, “Managing a large database of array interpolated images,” IEEE Trans. Signal Process., vol. 53, no. camera fingerprints,” in Proc. SPIE Conf. Media Forensics and Secu- 10, pp. 3948–3959, Oct. 2005. rity, San Jose, CA, 2010. [8] M. Kirchner, “Efficient estimation of CFA pattern configuration in dig- [35] S. Bayram, H. T. Sencar, and N. Memon, “Efficient techniques for ital camera images,” in Proc. SPIE Conf. Media Forensics and Secu- sensor fingerprint matching in large image and video databases,” in rity, San Jose, CA, 2010. Proc. SPIE Conf. Media Forensics and Security, San Jose, CA, 2010.
    • KEE et al.: DIGITAL IMAGE AUTHENTICATION FROM JPEG HEADERS 1075 Eric Kee received the B.S. and M.S. degrees in Hany Farid received his undergraduate degree in computer engineering from Virginia Tech in 2001 computer science and applied mathematics from and 2003, respectively. He is currently working the University of Rochester in 1989, and the Ph.D. toward the Ph.D. degree in computer science at degree in computer science from the University of Dartmouth College, Hanover, NH. Pennsylvania in 1997. He worked at Johns Hopkins University, Applied Following a two year postdoctoral position in Physics Laboratory until 2007. Brain and Cognitive Sciences at MIT, he joined the Computer Science Department at Dartmouth in 1999. He is the William H. Neukom 1964 Distin- guished Professor of Computational Science, and the Director of the Neukom Institute for Computational Science at Dartmouth College. Dr. Farid is the recipient of an NSF CAREER award, a Sloan Fellowship, and a Guggenheim Fellowship. Micah K. Johnson holds undergraduate degrees in mathematics and music theory from the University of New Hampshire, the A.M. degree in electro-acoustic music from Dartmouth College, and the Ph.D. degree in computer science from Dartmouth College. He is a Research Scientist at MIT. His research in- terests span computer vision and graphics, with an emphasis on image analysis for shape, illumination, color, texture, and realism.