Appendix A - Glossary

🌐 中文版

Arranged roughly in reading order. Mastering the bold-faced terms covers almost everything in this handbook.

Term	Also known as	One-line meaning
Steganography	-	Hiding a secret inside an innocent carrier so the act of communication is hidden
Steganalysis	-	Deciding whether a carrier hides a secret, from statistics or learning
cover / stego	carrier / embedded image	Original image / image after embedding
LSB	least significant bit	Lowest bit of a pixel; flipping it is visually negligible
bit plane	-	Binary layer formed by one bit position across all pixels
redundancy	-	Carrier parts that can be modified without obvious artifacts
capacity / payload	embedding capacity	Maximum secret bits an image can carry
relative payload	-	Embedded bits divided by available positions
embedding efficiency	alpha	Average secret bits per modification, alpha
matrix embedding	-	Group coding that embeds p bits while changing at most one position
Hamming code	binary Hamming code	Linear error-locating code with block [n,k,3]
parity-check matrix	H matrix	p x n matrix with distinct columns used to compute syndromes
syndrome	-	s = H*x (mod 2); a summary of the block state
GF(2)	Galois field of two elements	Field containing only 0/1 with XOR arithmetic
shrinkage	-	Coefficient magnitude decrease lands on zero, ruining the block
wet paper coding	WPC	Embedding on dry positions while wet positions stay untouched
dry / wet positions	-	Modifiable positions / untouchable positions
F5 / nsF5	-	JPEG matrix-embedding algorithm; nsF5 removes shrinkage with wet paper
magnitude decrease	coefficient shrink	Reducing |coefficient| by 1 to flip LSB parity
SHA-256	-	256-bit cryptographic hash used for image keying
keying	-	Using password/content keys to decide hiding positions
deterministic permutation	-	Shuffle where the same seed always gives the same order
splitmix64	-	64-bit PRNG used by the project
Fisher-Yates	-	Standard uniform shuffle algorithm
self-synchronization	-	Decoder finds the payload without external parameters
tamper perception	-	Detecting that an image was modified
blind steganalysis	-	Statistical detection without the original cover
chi-square test	Westfeld test	Checks whether gray-level pairs were balanced
p-value	-	Probability that observations fit the assumption; here high p is suspicious
RS analysis	-	Counts regular/singular groups under +/- masks
regular / singular	R / S groups	Groups whose discriminant rises / falls after flipping
difference entropy	-	Shannon entropy of adjacent differences; measures texture randomness
Shannon entropy	-	Information/uncertainty measure in bits
supervised learning	-	Training a model on labeled samples
feature vector	-	Numeric description of a sample (11-D v1 or 143-D v2 here)
label	-	Ground truth (0 = clean, 1 = stego)
binary classification	-	Predicting membership of one of two classes
logistic regression	-	Linear combination + sigmoid; an interpretable classifier
sigmoid	-	S-shaped function mapping any real number to 0-1
cross-entropy loss	-	Standard classification loss
gradient descent	-	Iteratively moving parameters downhill in loss
overfitting	-	Memorizing training data at the cost of generalization
cross-validation	-	Rotating validation folds to estimate generalization
data leakage	-	Validation information entering training, inflating metrics
GroupKFold	-	Grouped cross-validation (same photo’s samples stay together)
confusion matrix	-	Four-cell counts of truth vs prediction
precision / recall	-	Share of flagged that are real / share of real that are caught
ROC / AUC	-	Detection-vs-FP curve across thresholds and its area
Youden’s J	-	Threshold maximizing TPR - FPR
false positive / negative	FP / FN	Clean flagged as stego / stego missed
SRM	Spatial Rich Model	High-pass residual filter family that highlights embedding noise
residual map	-	Noise residual after high-pass filtering
LightGBM / LGB	-	Gradient-boosting implementation used by the v1.4 models
feature group	-	A batch of features from one source (BASE/SRM/PREFIX/LSB-PREFIX)
out-of-distribution	OOD	Inputs unlike the training distribution (e.g., real JPEG clean)
BOSSbase	BOSSbase 1.01	Standard benchmark: 10,000 512x512 grayscale PGM images
PGM	-	Lossless grayscale image format (BOSSbase carrier format)
stacking	meta-learner	Ensemble where a second model combines base-model predictions
memmap	memory-mapped file	Large arrays written/read in slices to avoid OOM
calibration set	-	Separate subset used only for threshold selection
grid tuning	-	Searching a hyperparameter grid and comparing models
dual-version model	-	143d robust model + 53d interpretable model deployed together
ctypes	-	Python binding mechanism for C/C++ DLLs
DLL	dynamic-link library	Windows shared library
CUDA / batching	-	GPU parallelism / processing many samples at once
consistency check	-	Bit-level comparison between C++/GPU and Python outputs