Day 3 of 1060 minShell WalkerHands-on lab
Day 3: Spotting AI-Generated Images
Taught by Mr Elite Updated
By the end of today you can
- Check an image for visual tells and rate how strong each one is
- Run error level analysis and a noise residual check with a short Python script
- Treat detector scores as leads and explain why they give false positives
- Reveal near-invisible hidden text and trailing data in an image
Warm-up
Answer from memory first, then tap to check. Recalling yesterday is what makes it stick.
Why is a clean EXIF block not proof that a photo is genuine?
Metadata is a set of claims inside the file and is trivial to forge, copy or strip, so presence and absence are both inconclusive.
Which command shows the real file type regardless of the extension?
file, or xxd to read the header bytes directly.
Someone sends you a photo of a 'burned out warehouse' that is spreading on social media. The EXIF is empty, the hands look slightly odd, and an online detector says 87 percent AI. Your manager wants a yes or no by lunch.
Today I teach you to stop asking for a yes or no from one signal. You will collect several small signals, rate how much each one is worth, and write down a confidence level you can defend. We also look at a second question that matters just as much: what is hiding inside the image.
One signal never decides
On Day 2 you learned that metadata is a claim, not a fact. Image analysis works the same way. Every technique today produces a signal, and every signal can be wrong. Your job is to collect several, rate each one, and see whether they agree.
I use four ratings. Strong means hard to fake or explain away. Moderate means suggestive and needs support. Weak means it happens in genuine images all the time. None means you looked and found nothing, which is not the same as proving the image is real.
- Artefact
- A flaw or trace left behind by how an image was made or processed.
- False positive
- A tool says 'fake' about a genuine image.
- False negative
- A tool says 'genuine' about a fake image.
Visual tells: look before you run anything
Open the image at full size and zoom in. Do this on your working copy, never the original. Generators are better than they were, but they still slip in four places.
Quick check
A sign in the background has slightly garbled letters, but everything else looks consistent. How should you rate this signal on its own?
Garbled text is suggestive but genuine low-quality photos produce it too, so it needs support from other signals.
Watch me do it: ELA and noise residual
Error level analysis (ELA) re-saves a JPEG at a known quality and shows how much each area changed. Regions that were edited or pasted in often respond differently from the rest. A noise residual subtracts a blurred copy to expose the fine grain a camera sensor leaves. Everything here runs on a file on your own machine.
mkdir -p work && cp ~/Pictures/photo.jpg work/photo.jpg
sha256sum ~/Pictures/photo.jpg work/photo.jpgimport io, os, sys
import numpy as np
from PIL import Image, ImageChops, ImageFilter
path = sys.argv[1]
img = Image.open(path).convert("RGB")
buf = io.BytesIO()
img.save(buf, "JPEG", quality=90)
buf.seek(0)
resaved = Image.open(buf).convert("RGB")
diff = np.asarray(ImageChops.difference(img, resaved)).astype(float)
print("ELA mean", round(diff.mean(), 2), "max", diff.max())
scale = 255.0 / max(diff.max(), 1)
out = Image.fromarray((diff * scale).astype("uint8"))
out.save(path + ".ela.png")
blur = img.filter(ImageFilter.GaussianBlur(2))
res = np.asarray(img).astype(float) - np.asarray(blur).astype(float)
print("Noise std", round(res.std(), 2))
noise = np.clip(np.abs(res) * 8, 0, 255).astype("uint8")
Image.fromarray(noise).save(path + ".noise.png")What just happened: the script amplified tiny differences so your eyes can compare regions. A bright edge around a face, or one smooth patch in a grainy photo, is worth a note. But screenshots, heavy filters and repeated re-uploads also create odd ELA, so I rate ELA as weak to moderate. Fully generated images often show unusually smooth, even noise, which is another moderate signal at best.
Quick check
ELA shows a bright outline around a person in a photo that has been shared on messaging apps several times. What is the sensible reading?
Recompression leaves its own traces, so ELA is a lead that needs corroboration.
Detector tools and their false positives
Online AI-image detectors give you a percentage. It feels scientific, but it is a classifier's guess trained on particular generators. A new model, a screenshot, a resize or a heavy filter can fool it in either direction.
Use a detector like a witness who has not seen the whole scene. Record the tool name, the date, the score, and that you ran it on a copy. Never upload sensitive or confidential evidence to a public web tool. Check your organisation's rules first.
Checking an image for hidden content
Images can carry things you cannot see: text in near-white on white, data appended after the end of the file, or odd strings in metadata. This matters for fraud and, as you will see in the lab, for AI systems that read images.
exiftool -warning -a -G1 work/photo.jpg
strings -n 12 work/photo.jpg | head -n 20import sys
import numpy as np
from PIL import Image
img = Image.open(sys.argv[1]).convert("L")
a = np.asarray(img).astype(float)
low = 235
stretched = (np.clip(a, low, 255) - low) * (255 / (255 - low))
Image.fromarray(stretched.astype("uint8")).save(sys.argv[1] + ".reveal.png")
print("saved", sys.argv[1] + ".reveal.png")Your lab: text a human cannot see
The lab below shows why hidden content matters. Approach it as a defender learning the risk. You will work only inside the SecurityElites lab, never against a real system.
Your turn
AI Visual Prompt Injection
ReceiptLens is a practice app that reads receipt images and writes a summary for a human approver. A vision model reads every pixel, including text a person would never notice. We use it now because it flips today's question: instead of hunting for hidden content in a suspect image, you see how hidden content can steer an automated reader.
What to try
What to notice
- Whether the summary reflects text you cannot see when viewing the image normally
- How little contrast difference the model needs to read text
- Which of today's checks (reveal stretch, strings, metadata) would have caught your hidden line
If it did not work
Treating a detector score as a verdict
Record it as one moderate signal with tool, date and version, then look for agreeing or conflicting evidence.
Running ELA on a screenshot or heavily shared image and reading the result literally
Note the image's history first. If it has been recompressed many times, rate ELA as weak.
Analysing the original file
Hash it, copy it, and work only on the copy, as you did on Day 1 and Day 2.
Writing 'no signals found, so it is real'
Write 'no signals found, confidence in authenticity limited'. Absence of artefacts proves little.
Recap
- Every image signal is a lead, rated strong, moderate, weak or none
- Visual tells are useful but modern generators are improving on hands and text
- ELA and noise residuals highlight inconsistencies, but compression creates false alarms
- Detector scores are classifier guesses, so record them and never rely on them alone
- Check images for hidden text and trailing data, because machine readers see what people miss
- Confidence statements should be stated and justified, not implied
Cheat sheet
python3 ela.py img.jpg- Writes ELA and noise residual PNGs and prints summary numbers
python3 reveal.py img.png- Stretches near-white values to expose faint hidden text
exiftool -warning -a -G1 file- Shows parser warnings such as trailing data
strings -n 12 file- Prints readable text runs of 12 or more characters
Rating scale- Strong, moderate, weak, none
False positive / negative- Genuine flagged as fake / fake passed as genuine
Before the next day
Mission: Five-image signal table
Examine five images, some genuine and some generated, and record the signals and a confidence rating for each. This is the core of Day 4, so keep your notes tidy. Use only images you own, generated with a free tool yourself, or that are clearly licensed for use.
About 35 min
You are done when: A markdown file with one table row per image: hash, signals found with ratings, overall confidence, and the final answer-key result.
Want more? Take one genuine photo, resize and re-save it three times, and note how ELA and detector scores change.
Quiz
4 questions. Your score is saved with your progress.
1. Three weak signals point towards generation and one strong signal points towards genuine. What is the best conclusion?
Signals are weighted by strength, not counted, so the strong signal carries more weight.
2. Why can ELA flag a genuine image?
Compression history changes error levels across an image, which can look like editing.
3. A detector gives an image 12 percent AI. What does this tell you?
A low score is not proof, because detectors miss new generators and altered images.
4. Why does hidden near-white text in a receipt image matter for defenders?
Automated readers process every pixel, so hidden instructions can influence their output without the human noticing.
Finish the day
Mark the day complete to save your progress.
Next: Day 4 is an apply-it day: you will triage a set of suspect images using the Day 1 to 3 methods in one workflow. Today's signal table is the format you will reuse. Coming soon.
Your notes
Private to this device. Write what you want to remember from today.