Visual Mis/disinformation in Journalism and Public Communications: Current Verification Practices, Challenges, and Future Opportunities

Types of Issues and Corresponding Detection Methods

As photo and video editing technology becomes more widespread and sophisticated, developed society is increasingly being exposed to more imagery with unknown provenance and accuracy. Journalists, media outlets, and law enforcement often cannot establish the veracity of an image by simply probing its source, and now look to use passive “digital forensics” tools in order to know whether an image's contents can be trusted (Songcharoen, Bite, and Clay Citation2014; Farid Citation2018). This section presents an overview of types of manipulation and opportunities journalists have to identify and evaluate them.

Considering the above-noted issues concerning the media literacy required to investigate visual fakery, we present below five common types of image manipulations alongside a grid (see Table 1) for an example of each category within a journalistic context alongside the best approach(s) to detect it.

• Copy-move. A copy-move or cloning operation takes a region of pixels within an image, copies it, and moves this copy into another area within the image. Such was the case with a 1964 photo of Dr Martin Luther King Jr, which was edited using the copy-move technique to make it appear that MLK was flipping off the camera (see Figure 2). Copy-move operations are among the most common image forgery techniques, as to the human eye they uphold the internal consistency of an image. Copy-move operations are also among the most commonly detected manipulations. They are often identified by looking for JPEG compression inconsistencies, CFA interpolation inconsistencies, contrast inconsistencies, and are detected by most deep-learning based general classifiers. In a journalistic context, peak bodies, such as the Associated Press in the US or the Media Entertainment & Arts Alliance in Australia provide guidance to their members about ethical conduct related to editing and presenting news visuals. Relevant to both copy-move and splicing operations is guidance from both organisations that states: “No element should be digitally added to or subtracted from any photograph” (Associated Press Citation2020) and “Present pictures and sound which are true and accurate. Any manipulation likely to mislead should be disclosed” (MEAA Citation2020). The US-based National Press Photographers Association encourages its members to “maintain the integrity” of the visuals’ content and context and to “not manipulate images or add or alter sound in any way that can mislead viewers or misrepresent subjects” (National Press Photographers Association Citation2020).

• Splicing. Splicing, also known as image composition, is similar to copy-move. Cloned regions from one source are copied into another to form a composite image (see Figure 3). Splicing is similarly detected by identifying JPEG compression inconsistencies, CFA interpolation inconsistencies, contrast inconsistencies, and by using general classifiers.

• Resampling. Resampling is often used in the process of resizing, rotating, stretching, or skewing in order to create images with different pixel densities. In a journalistic context, the Associated Press's editing and presentation principles allow backgrounds to be removed and figures to be overlaid over a neutral background for graphics and stipulates “such compositions must not misrepresent the facts and must not result in an image that looks like a photograph—it must clearly be a graphic” (Associated Press Citation2020). Often during a copy-move or splicing operation, copied regions that don't have identical dimensions need to be resampled in order to create convincing forgeries. The algorithms used to resample images can leave artefacts and statistical differences within resampled regions. This is often detected by looking for a CFA interpolation inconsistencies, noise inconsistency, and by using general classifiers (Figure 4).

• Retouching. Retouching is the process of making small adjustments to emphasise or obfuscate features of an image, such as blemishes, acne, or scars. Retouching is often done as part of the post-production process for publishing photos of people, and recently has become popular among amateur photographers on social media. Such retouching has been done on photos of US President Donald J Trump that have been shared to his official Facebook and Instagram accounts (see Figure 5) in order to make him appear slimmer and with more flattering hair. Within a journalistic context, the only retouching allowed under Associated Press principles is “to eliminate dust on camera sensors and scratches on scanned negatives or scanned prints” (Associated Press Citation2020). Retouching is often detected by looking for CFA interpolation inconsistencies, noise inconsistency, and by using general classifiers (Figure 5).

• Cropping. Cropping removes unwanted areas at the edges of an image. Cropping is a standard, widespread, and “acceptable” form of news image editing (Zhang Citation2018) unless the intent is to deceive. It is often used in image forgery, however, to hide objects or conceal context. Such was the case in 2017 when, under pressure from the White House, the U.S. National Parks Service cropped official photos taken from the Washington Monument that showed the size of US President Donald J Trump's inauguration crowd size. The cropped photos provided less overall context and made the overall crowd size less apparent (Figure 6). Such cropped edges are often highlighted when looking for JPEG compression inconsistencies.

Figure 2. MLK comparison.

Note: In the original photo, at right, Dr Martin Luther King Jr reacts in St Augustine, Florida, after learning that the US Senate passed the civil rights bill on June 19, 1964. An altered version of the image, using the copy-move technique, clones part of the background over MLK’s second finger to make it appear that he is flipping off the camera. The edited image was shared widely on Twitter, Reddit, and on the white supremacist website Daily Stormer.

Display full size

Figure 3. Australian bushfire composite.

Note: The photo on the left, which was shared widely on social media during the 2019–20 Australian bushfire crisis, is a composite of at least three other images that have been digitally combined in Photoshop using the splicing technique.

Display full size

Figure 4. Hurricane Dorian composite.

Note: The composite image, at left, which uses multiple manipulation techniques, and was shared on social media with a caption that claimed it portrayed Hurricane Dorian swirling over Florida, used resampling techniques to make the horizontal photo of a storm cell in Kansas fit the vertical composition at left.

Display full size

Figure 5. Donald J. Trump comparison.

Note: The original photo of Donald J Trump (left) and the edited version of him (right) that he shared on his official Instagram and Facebook accounts and that shows him with elongated fingers, a tightened waistline and higher crotch, a slimmed neck and shoulder, and tightened hair.

Display full size

Both active and passive techniques can be used to verify the authenticity of an image. Active techniques like digital watermarking are able to identify individual photographers or camera equipment, as well as uphold the veracity of an image. Digital watermarking invisibly embeds authentication data into original images, which can later be derived to verify said image (Shih Citation2017). The most prominent obstacle to the widespread use of digital watermarking is that photographers and media organisations don't always have incentives to use it, and it places the burden of proof and cost upon the media creators who need to use additional third-party tools during post-production (Zhou and Lv Citation2011).

In contrast to active techniques, passive (sometimes called blind) techniques look for visual artefacts and anomalies in the underlying composition and derived statistics of images to identify image manipulation or forgery. These tools are powerful because they can be applied to images with unknown provenance, and identify an ever-expanding list of image manipulation techniques. They are limited however, as they cannot comprehensively “prove a negative”, and assessing the accuracy of these techniques can be challenging (Marshall and Paige Citation2018).

There are many existing literature reviews and surveys in the field of image forensics (Farid Citation2009; Birajdar and Mankar Citation2013; Qureshi and Deriche Citation2015), although few that consider newer deep-learning techniques, and none that do so in the context of journalism and public communication. These reviews exist primarily in the computer science and media forensics fields, and while not entirely divorced from the contexts that their evaluated techniques can be used in, their findings are intended to be broadly applicable and don't necessarily reflect pragmatic usage.

Passive Detection Methods

Passive statistical and algorithmic techniques look for manipulation artefacts and analyse the underlying statistics of an image. Although visual artefacts are not always recognisable or visible, the underlying statistics of images are changed when they are manipulated. Table 2 provides an overview of each approach alongside its strengths and weaknesses.

• JPEG compression inconsistency. The JPEG format is used to provide a universally accessible way to compress and display images with varying compression ratios (Marqués, Menezes, and Ruiz-Hidalgo Citation2009). JPEG sees widespread use online because it significantly reduces an image's file size with only minimal changes to its appearance. JPEGs’ discrete cosine transform (DCT) compression process is “lossy”, meaning that accurate pixel information is lost during compression. Splicing one JPEG image into another with photo editing software causes “double JPEG compression” (Popescu and Farid Citation2004), as the image is compressed a second time. This second round of compression introduces changes such as periodic artefacts in the DCT coefficient distribution of blocks within the image. This unusual periodicity can be detected using statistical tools (Fan and De Queiroz Citation2003). It should be noted, however, that double-jpeg compression is not in itself evidence of manipulation, as it could also be caused by incidental compression of an already compressed image. This is common in scenarios where for example, a photo is taken on a camera and saved as a JPEG, and then uploaded to a social media service where it is compressed again to decrease file size. This makes its universal usage not ideal for the contexts that this study is considering, as it would have a high false-positive rate. It would however be a useful tool in cases where a JPEG file is deceptively presented as having been unaltered from its original state, such as in visual journalism competitions, like World Press Photo, where the amount of post-production can be limited if allowed at all, or in the earlier example (see Figure 6) of the US National Parks Service cropping official photos of US President Donald Trump's 2017 inauguration. Fu, Shi, and Su (Citation2007) demonstrate a way to determine whether an image has been JPEG compressed, and to estimate its compression quality (1–100), which can give clues about its provenance. Farid (Citation2006, Citation2008) also found that the way that a JPEG's quantisation table (part of the compression process and reflected in the file's headers) is indicative of the make and model of camera used to create the image, which can be used for authentication.

• CFA interpolation inconsistency. In digital photography, images are created by capturing light in arrays of small cells known as photosites or photoreceptors. When a camera's shutter is open for a given exposure time, each cell captures and quantifies an amount of light which is then digitised. This value in each of the cells is combined and used to create an image. In order to include colours in images, varying filters are applied to each photosite that will block out unwanted light, often such that only one colour is quantified per cell (Malpas Citation2007). This arrangement of cells is known as the Colour Filter Array. Because only one colour is quantified per cell, neighbouring cells are used to create a pixel with multiple colour channels through estimation algorithms by sampling surrounding cells, a process known as interpolation. As noted by Popescu and Farid (Citation2005), the various algorithms used for interpolation introduce statistical correlations within images that are disrupted or made inconsistent when an image is forged, for example by copy-move or splicing operations. They demonstrate that statistical inconsistencies caused by tampered regions can be detected and thus locally identified. Goljan and Fridrich (Citation2015) similarly note that there is usually a stronger correlation between colour channels than merely among neighbouring pixels, which can be analysed to detect manipulations. Such statistical inconsistencies can also reveal artificial blurring of images, such as when the US National Archives in 2020 blurred images critical of US President Donald Trump (see Figure 7). The archives did not disclose the edit nor did it respond to requests for a list of other edits it made “so as not to engage in current political controversy”, which was the rationale offered for the change.

• Contrast and lighting inconsistency. An image's contrast is the range of difference in brightness or colour between its elements. This is sometimes edited in extreme ways for political effect. A historic example is the darkening of OJ Simpson's mugshot on the 1994 cover of TIME magazine and a more contemporary example is a photo of US Donald Trump (see Figure 8) that was posted on Twitter and appears to have increased the saturation and/or contrast of the president's face, which accentuated the difference between tanned and untanned parts of his skin. The Tweet attracted widespread attention and even drew a mention by Trump himself who claimed it was Photoshopped.

  Adjusting contrast is also often changed when attempting to splice two images with varying contrasts together in order to appear genuine (see Figure 4). Forgeries that don't change contrast can often be detected by the human eye and can also be detected using statistical tools. Lin, Li, and Hu (Citation2013) demonstrate that adjusting an image's contrast is detectable by analysing the correlation between colour channels as discussed in CFA interpolation inconsistency. Stamm and Liu (Citation2010) also demonstrate that contrast enhancement can be detected by analysing the “smoothness” of the frequency of change in colour across images (which appears more uniform in genuine photographs), and furthermore, that the degree of contrast enhancement can be derived through an iterative algorithm. Image light sources can be hard to manipulate in image forgery as they can be difficult to derive with the human eye. Johnson and Farid (Citation2005, Citation2007) show that when correctly derived with computer vision techniques, splicing operations can be detected.

• Noise inconsistency. Noise is a common and, depending on lighting conditions and the camera's settings, reasonably invisible component of digital photographs that results from statistical variations in the camera sensor being translated into visual artefacts. As noise is often (but not always) uniform in genuine photographs, image forgery can be detected by looking for inconsistencies in noise across the image (see Figure 7). Pan, Zhang, and Lyu (Citation2012) demonstrate this by estimating global noise variance across the image and thus determining if it has been manipulated. Chen et al. (Citation2008) additionally show that an image's origin can be verified by analysing the noise on the image and comparing it to the noise expected from a given sensor.

Figure 6. Trump inauguration comparison.

Note: A US National Parks Service photographer edited official pictures of Donald Trump’s inauguration to make the crowd appear bigger following a personal intervention from the president, according to documents obtained through a Freedom of Information Act request lodged by The Guardian in 2018. The NPS photographer cropped out empty space “where the crowd ended” for a new set of pictures requested by Trump on the first morning of his presidency, after he was angered by images showing his audience was smaller than Barack Obama’s in 2009. Such edits could have been revealed through an analysis of the JPEG file’s compression details.

Display full size

Figure 7. Women's March edits.

Note: The US National Archives sparked controversy in 2020 when it edited a news photo by Mario Tama for Getty (pictured) of the 2017 Women’s March by blurring out certain words, such as a reference to US President Donald Trump in a sign that read, “God Hates Trump”, among other edits.

Display full size

Figure 8. Trump skin comparison.

Note: A photo, at left, which depicted an extreme difference between the tanned and untanned portions of US President Donald Trump’s face, attracted widespread attention and drew condemnation from Trump himself. Other photos taken the same day, such as the one at right, by Al Drago for Bloomberg, show a similar but less extreme difference. Passive forensic methods, such as evaluating the image for contrast/lighting inconsistency, could provide certainty to the claims that the image at left was Photoshopped.

Display full size

Deep Learning Methods

Deep learning for the identification of image manipulation is growing in popularity, largely due to the boom in the generalised use of machine learning and artificial intelligence. Deep learning refers to a subclass of neural network algorithms (deep neural network, or DNN) that can be trained to classify input information (in this case pictures) into specific predefined categories, such as photos of Donald Trump versus photos of Justin Trudeau, for example. As opposed to traditional forensic algorithms (ie, the “passive” methods detailed above), DNNs don't require prior understanding of precisely how images are manipulated. Instead, they require a representative sample of manipulated and unmanipulated images (training data). These training data are fed iteratively through the network and a learning algorithm is used to modify a vast array of internal parameters to gradually maximize classification accuracy on these samples. Training finishes when the classification accuracy remains unchanged, known as convergence, indicating that the DNN is now “trained”. The assumption behind the “train-by-example” approach is that, if the training examples represent a comprehensive enough sample of images of a phenomenon, then an adequately resourced and configured DNN can create an abstract internal representation of what image qualities combine to indicate manipulation (Patterson and Gibson Citation2017). This abstracted internal model means that the trained DNN can then confidently identify manipulated images in a generalised way (Rao and Ni Citation2016). A trained model can therefore identify both the common artefacts that are left behind after image manipulation operations and underlying statistical anomalies.

The most common and accessible DNN architecture for image classification is the convolutional neural network (CNN). Outside digital forensics, CNNs are typically used to identify and classify discrete objects in a given image (such as in, for example, Facebook's facial recognition photo tagging functionality). In digital forensics, CNNs can similarly be used to identify the manipulated regions of an image (Zhou et al. Citation2018) (see Figure 9). However as Chen et al. (Citation2015) found, generic CNN models are heavily affected by object edges and textures, and may not be suited to identify small image manipulation artefacts, such as airbrushing out a zit or enlarged pore in contrast to adding or removing an entire discrete object.

Figure 9. Mask R-CNN example.

Note: This image, taken from a standardised image library designed to train and test machine vision classifiers in the detection of image manipulation (Silva et al. Citation2015), demonstrates how a deep learning approach might work in practice. “Masking” (the area shaded in red) is a technique used in machine vision that seeks to separate (segment) different objects in an image or a video. Given an input image, a masking algorithm provides object bounding boxes, labels, and a mask. In the case of image manipulation, these masks could be used to highlight regions of suspected pixel manipulation, the form of manipulation believed to have taken place, and a confidence estimate. In this example, the algorithm has identified roughly 20 percent of the image as potentially having been forged, has labelled the type of potential forgery as a copy-move manipulation, and estimates with 92 percent confidence that the forgery has taken place.

Display full size

Due to their generalised approach, under certain circumstances deep-learning-based general image manipulation detection tools may demonstrate an advantage compared to specific image manipulation detection tools. For example, rather than a journalist having to perform multiple individual tests to identify whether the image had been cropped, selectively blurred, or made more or less saturated, a single deep learning method could, in theory, identify all three of those edits as well as other ones that might not be apparent (such as retouching, resampling, splicing, or copy-moving). And in the contexts that this review considers, they can present a clear usability advantage in that there is usually a single (binary, scalar, or visual) output. Additionally, because a single tool can be used to investigate a variety of manipulations, it is simpler to account for false positives and negatives than when using an assortment of single-purpose tools. There are, however, clear disadvantages to this approach. From the user's perspective, the results can be visual and localised so the viewer can helpfully see which region or regions of the image are problematic; however, the underlying logic related to why certain regions are classified the way they are isn't transparent. The way that data are sourced to train deep-learning based classifiers is also problematic given the relatively small size of popular training datasets and the limited availability of datasets demonstrating the diverse range of sophisticated image manipulation techniques, as well as adversarialFootnote³ ones.

The adversarial issue is most pertinent though, as following the development of forensic methods that establish the trustworthiness of images, anti-forensic methods have also been developed expressly to thwart traditional forensic methods. These adversarial methods present critical challenges when using image-manipulation tools practically. Anti-forensic methods rely on knowing how particular forensic methods work, either through reverse engineering or using available knowledge. It is then possible to derive and test adversarial examples which modify images such that they won't be flagged by classifiers.

Stamm et al. (Citation2010) and Stamm and Liu (Citation2011) demonstrate that methods detecting JPEG compression inconsistencies can be fooled by modifying the intrinsic statistical properties of images in ways that allow them to appear both visually and statistically ordinary. This can be done by adding precise dither (i.e., noise) to the underlying DCT coefficients, such that their distribution (as discussed in the JPEG compression inconsistency section) doesn't appear irregular or blocky. This highly effective technique severely undermines the efficacy of JPEG-compression based forgery identifiers and analysers.

Anti-forensic tools are also effective when used against deep-learning forgery detection methods. Szegedy et al. (Citation2013) found that by making imperceptibly small adjustments to the pixel-contents of an image, deep neural network-based classification models can be misled. And as Nguyen, Yosinski, and Clune (Citation2015) found, DNNs can even be fooled into misclassifying images that are visually meaningless, examples including images that appear to a human eye as noise, that are classified by a DNN as objects such as animals, fruit and household objects. While these techniques describe adversarial examples for traditional DNNs, the principals are similarly applied to forgery classifiers.

Even though developing and using anti-forensic techniques is non-trivial for a forger, the methods’ considerable efficacy and versatility demonstrate the extreme importance for consideration to be given to them when developing new classification tools. Furthermore, in practical contexts, the fallibility and limitations of image forgery classifiers needs to be communicated to journalists in order to give them realistic expectations and mitigate so-called “automation bias”. An ongoing commitment to the development of such tools is also essential for journalists and the public so that the tools remain relevant in the “digital arms race” of using computational approaches both to verify and frustrate verification efforts.

Implications for Journalistic Practice

Considering the wide variance of content that can fall on Wardle’s (Citation2018) “information disorder” spectrum (everything from satire and false connection to false context and fabricated content), the firehose of content journalists confront on a day-to-day basis, and the intense time pressures they have, especially on social media where breaking news unfolds, journalists face serious challenges when it comes to verifying visuals online. It's disheartening to recall past research that has found that only 11 percent of journalists globally use social media verification tools and to consider how much content of unknown provenance or authenticity they might be complicit in sharing, publishing, and amplifying.

Some tools, such as TinEye (a reverse image search engine), Foto Forensics, or Izitru, exist, but they would have to be used in combination and journalists have little time to manually run an image through each tool, compare the results, and decide. Additionally, reverse image search engines are not foolproof solutions, as major options like Google don't crawl the entirety of the web, such as images on 4chan or Reddit, rely on original and unedited versions of images being available, and can also be fooled by relatively simple edits, such as flipping the image's orientation from left to right. As such, without abdicating their verification duties entirely, journalists need to outsource some of the initial verification labour to machines that are more efficient with automated tasks to quickly run a battery of tests and return a decision—with rationale—to the journalist for final vetting and verification.

As an example: when US President Donald Trump posted on his official Facebook and Instagram accounts in 2018 and 2019 photos of himself that had been digitally edited to make him appear slimmer and to change the length of his fingers (Novak Citation2019), a statement like, “There is an inconsistency in the way that the smooth hue transition around the hands and waist, respectively, has been interpolated, which indicates a forgery in these regions of the image”, which is clear and falsifiable, is more useful in a journalistic context than “Our model says that there's an 88 percent certainty that this region of the image is forged”. Machines could process large amounts of data on journalists’ behalf and raise red flags for journalists to further investigate and decide on. By using such a funnel approach, journalists are able to dedicate their precious time to the most suspect of cases and machines are able to save journalists the time and labour of winnowing from the field various low-risk images. These efforts don't have to be siloed and duplicative, however. The experimental and prototypical DejaVu tool (Matatov et al. Citation2018), which isn't yet publicly available, proposes a database approach to allow journalists to flag problematic images and to be notified via a browser-based extension when such problematic images surface on their screens.

Vetting processes would need to happen at two critical junctures in the news gathering and dissemination process: first, at the point of discovery, such as when a journalist is trawling social media and considering amplifying a post by sharing it, and second, at the point of distribution/re-distribution on the associated news organisation's website, such as when a journalist embeds a Tweet into a news article or when a journalist uploads media into the organisation's content management system (CMS). This would provide verification opportunities both for user-generated content on social media as well as content that freelancers or staffers produce as part of their official duties.

Regarding the point of discovery, room for additional public-private partnerships exists. Some of the biggest tech companies are the ones best positioned to tackle these issues as they have the largest training datasets available and the technical infrastructure and engineering teams to handle a tool, ideally co-designed among developers, journalists, and members of the public, that could be integrated directly into social media platforms that would allow a user to, for example, click a button next to a post and have a report generated in return that provides an evaluation of the content's provenance and veracity. However, it is important to note that platform companies that integrate news into their offerings, such as Google, Facebook, Yahoo, and Twitter, are disproportionately benefiting from the advertising revenue that the more engaging mis/disinformation brings to their companies (Bakir and McStay Citation2018) and so would have to balance the potential revenue loss against the fear of tighter government regulation or fines for allowing mis/disinformation to flourish on their platforms.