by Blake Sawyer, Director of U.S. Operations, Amped Software
How often do we trust what we see on the internet? In the past few months, we have gone from obviously fake-looking AI-generated images to images and even videos that are becoming increasingly realistic. Add to that, many of the creative tools have low costs of entry (e.g., OpenAI’s Sora v2 is free if you can get an invitation). This accessibility contributes to a perfect storm for initial distrust of images and videos that are viewed on social media, shared with friends, and co-worked, or even submitted as “proof” of a crime. Recently, there was even a “prank challenge” of a homeless person breaking into a house, which has led to an increase in emergency calls and warnings from law enforcement to the public not to believe everything they see.
However, looking for clues and “interrogating” media (images, video, and audio) is not a new practice in investigations. Not only that, but the same principles apply when investigating any crime.
When a witness is interviewed to give a record of events, what steps do officers take to validate that testimony?
When a crime is committed, and a weapon is found or used, what are the standard operating procedures for retrieving and keeping that evidence as preserved as possible?
When a video or image is submitted as evidence, what steps are taken to ensure the information is accurate and unaltered?
Depending on your agency, there are processes in place to investigate and preserve the integrity of the first two questions, but fewer precautions are likely in place for the third. Even still, the presence of image and video evidence as primary evidence continues to grow exponentially. As this evidence accumulates in volume, its ubiquity renders direct acquisition a substantial drain on agency resources. This inevitably results in pictures and videos being submitted by witnesses, victims, or citizens, which are then recaptured from online sources.
In speaking with the U.S. Senate Subcommittee on Consumer Protection, Product Safety and Data Security, Sam Gregory (Executive Director for a nonprofit human rights group WITNESS) stated “Significant evolutions in volume, ease of access, personalization and malicious usage of generative AI reflect both the potential for creativity but also the heightened harms from audiovisual generative AI and deepfakes–including the plausible deniability that these tools enable, undermining consumers’ trust in the information ecosystem.” As such, law enforcement’s ability to assess and validate the media they receive is becoming a vital step in investigations.
Submitted media and potential pitfalls
Over the last several years, there has been a growth in the number of options for the public to submit images and videos directly to agencies. This has been an enormous help, as it enables officers to review the evidence more quickly and disseminate pertinent information to the public. This also frees resources for the agency, as an officer or video examiner no longer needs to go “into the field” to acquire a file that a witness or victim can easily transfer. Depending on the file's length, this can save hours at the scene, allowing investigators to return to casework and examiners to maintain a low backlog.
The challenges arise when questions of originality, authenticity, and file integrity are raised. Additionally, the submitter may be required to testify to the file's origin, which can be difficult in circumstances where the submitter is sending a file under the condition of anonymity. As seen in multiple examples, a file can be altered either innocently or with deceptive intent, and these changes may impact the admissibility of the file. Changes may be significant, such as adding or removing objects or persons from the scene, or as small as changing the date/time of a file’s creation in its metadata.
In a recent example, a video file was submitted to the prosecuting attorney’s office several days prior to the trial's scheduled start. The attorney received the file via Airdrop, but when the defense attorney didn’t have an Apple device, they shared it via text message. As a result, the original was not saved, and both parties worked with different copies of the evidence without realizing they were referring to different files until the case was in deliberation. This simple process may have affected the course of the trial if not remedied.
What does authentication mean?
When discussing image and video authentication, various terms are often combined. For example, when a court asks if a video is authentic, it could mean:
“Is the video the same as when it was submitted?”
“Is the video a clear and accurate depiction of what it purports?”
“Is this an original recording, or a copy?”
“Are there signs of tampering or manipulation?”
When examining a digital image or video, it is essential to separate and address each question individually. In Amped Software’s Authenticate training, these questions are grouped into categories: source identification, context analysis, integrity verification, and processing and tampering analysis.
Source identification
When a file is either acquired or submitted as evidence, one of the first questions to be answered is, “Where did this file come from?”. Source Identification assesses at which level an examiner can validate where an image or video originated. These can mean answering questions like “What type of device recorded this?”, “What was the make and model of the camera that recorded this?”, “What specific device was this recorded on?”, and even “Was this file saved from or sent across the internet?”.
For example, the image below was taken with an iPhone 15 Pro Max. In the source identification, this would answer two of the questions. By digging deeper into the file, we can, perhaps, tie it to a specific iPhone so we can know whether the device that submitted it recorded the image or was sent to that device. If the file were sent to the device, we would want to know more about how it was sent (e.g., a cloud service, a text message, or saved from social media). We can also look for “clues” that it was opened, edited, or saved using a software application.
Context analysis
Another common request is whether the media accurately and fairly depicts what it claims to be. A vital part of answering this question involves context analysis, which addresses what is in a photo or video and the circumstances around the image. Often, this involves examining the context of the scene, including the time and geographical context, as well as the process for creating digital files and publishing.
This category is helpful because it helps address images where scenes from a protest are reused from previous events, where the time and dates within the image do not match the time and dates of the incident, or where the events were staged as a reenactment. One recent example received through Amped Software’s support came as part of an alibi in a homicide case. The suspect in this case had submitted a video from social media that placed the suspect in another state at the time of the murder. In the context of this video, the dates and time indicated a recently recorded video, with no data matched or placed the video in the reported state.
Another example where context is in question would be a video posted to social media. A further examination would be needed to ensure that the entirety of the video was submitted, rather than whether it was trimmed before being posted. This can be done for several reasons, including innocent ones, such as a social media platform only allowing a maximum length for videos or cropping images to fit a square format. In the image below, we can see the footage from a bombing posted to a news channel’s feed. It shows the same explosions repeated three times so that the video fills the preset timing of a video clip and, perhaps, makes the event appear more dramatic.
Integrity verification
Every image or video submitted to a department should have precautions to ensure the media file is unaltered from the time of acquisition. Steps commonly used for digital files include documenting the chain of custody, writing the file to a storage medium that cannot be altered (e.g., DVD-R and Blu-ray-R), and generating a digital fingerprint for the file (a process known as “hashing”).
But what happened to the file before it was given to the examiner? Integrity verification is the process of examining other files that follow the same source generation, as well as clues left in the structure or metadata, to identify whether inconsistencies may exist. This involves examining how most files on a device are created and identifying any inconsistencies in the evidence item. For example, suppose many images are saved on a phone in .jpg format, but the one submitted from the incident is saved in .png format. In that case, this may indicate that the image didn’t originate from the device or was screenshotted or resaved at a later time. In a recent test, a video saved in one format (H.265) on the phone was changed to an entirely different format with more compression (H.264) just by being sent through a cloud portal.
When a file cannot be compared to others from the same device, we can look for additional clues based on the standard image generation model. The image below appears to be consistent with the expected output of the camera that took it in several free public tools. However, comparing it to a common generation model reveals several inconsistencies that may indicate this was edited, cropped, or resaved before submission.
Processing and tampering analysis
While this is the most commonly thought of analysis when we ask about image and video authentication, it is often not the first one investigated. While a visual analysis can often reveal blatant attempts at tampering (for example, an edge that doesn't match in a swapped face or AI-generated text in the background), as technology improves, the human eye alone may be insufficient to detect the changes.
Images have been subject to tampering and processing almost since the inception of cameras. In ”DEEPFAKES IN THE DOCK: Preparing International Justice for Generative AI,” Raquel Vazquez Llorente reports that “spirit photography” has been documented as far back as 1869. Since the early days of silent film, special effects teams have been pushing the boundaries of what is possible for creative uses in movies and films. And much like today, tools and experts have been needed to “prove” what is real and identify signs of manipulation.
Generally, these processes range from cropping an image or video to adding or removing details, or even replacing a subject's face in an image or video. As technology advances, new tools and methods are needed to identify signs of tampering. Examples of these techniques include looking for signs of double compression, cloned or duplicated areas, and even geometric inconsistencies (such as shadows). These processes, which differ mathematically based on the file type, can help find unreliable regions in both images and videos.
As generative tools become more prevalent, new tools and techniques are needed to address them. Amped Software’s research indicates that current AI models consistently struggle to generate shadows from a single light source or align the perspective correctly of objects and reflections. We have also created advanced detection tools, such as the Face GAN Deepfake filter and the new Diffusion Model Deepfake filter, to aid examiners in determining if what they are seeing is a camera original or something that was artificially created. These techniques must be reliable for current models but also adaptable as technology evolves. This is why no single filter will find every sign of tampering, and also why Amped Authenticate employs several filters.
In this image, multiple filters are used to show that the face on the right was added later, and that it was created using a Deepfake method known as a GAN (Generative Adversarial Network).
Resources and best practices
These processes are vital to adding credibility to media that agencies receive or do not acquire directly from a source near the time of the incident. For agencies just beginning this process, it can be overwhelming. Thankfully, several resources are available to aid in image and video authentication. The Scientific Working Group on Digital Evidence (SWGDE) has best practices for both image and video authentication. Companies such as Adobe, Microsoft, and Google utilize a watermarking application called C2PA to help document when generative AI tools are employed. Lastly, Amped Software builds on both of these suggestions with an application called Amped Authenticate, which uses multiple filters and processes to examine images and videos for all four authentication aspects.
So, while bad actors have tried to obscure their actions in media files for many years, having the proper understanding, the right tools, and a healthy skepticism will help validate the information your agency receives in submitted media. The old saying “You can’t believe everything you see on TV” has now become “You can’t believe everything you are sent as evidence.”