🔮 Predicted Nov 2026 End-Sem Paper + Keyword Key

High-probability predicted paper (100 marks) for CBE412, based on the question-recycling pattern across the four PYQ papers, and factoring in the Sep 2026 Mid and Oct 2026 Repeat Mid. Keywords-only answer key.

PART A (4 × 5 = 20 Marks)

Q1. QoS Specification / Service Classes (Unit 4 — 5m)

• Prediction: A streaming/tele-medicine app needs guaranteed quality under fluctuating network conditions. Explain Application-level QoS specification parameters (Single, Pair, Triple deterministic bounds) OR compare Guaranteed, Predictable ($B^p$), and Best-Effort service classes (or Isochronous transmission for live conferencing).
• 🔑 Keywords:
  • 4 QoS Layers: Perceptual $\rightarrow$ Application $\rightarrow$ System $\rightarrow$ Network/Device.
  • App QoS Params: Frame rate (fps), Frame resolution (pixels), Color depth, Lip-sync skew ($<\!80\text{ ms}$).
  • Deterministic Bounds: Single (target), Pair $[QoS_{\min}, QoS_{\max}]$ (splits Acceptable vs. Unacceptable regions), Triple $(QoS_{\max}, QoS_{\text{avg}}, QoS_{\min})$.
  • Service Classes: Guaranteed (Resource Reservation, hard bounds), Predictable (historical sliding window $B^p = \frac{1}{L}\sum_{i=n-L}^n B_i$), Best-Effort (no guarantee).
  • Isochronous: Bounded end-to-end delay + minimal jitter + continuous synchronized flow.

Q2. Audio Sampling / Nyquist / Bit-Depth Quantization (Unit 2 — 5m)

• Prediction: Given an audio signal with max frequency $f_{\max}$ (e.g., $15\text{ kHz}$ or $750\text{ Hz}$), find minimum Nyquist rate, explain aliasing, why practical $f_s = 44.1\text{ kHz} > 2f_{\max}$, and how bit depth ($8\text{ bit} \rightarrow 16\text{ bit}$) affects quantization noise.
• 🔑 Keywords:
  • Nyquist Rate: $f_s \ge 2 f_{\max}$ (mathematically lossless sampling).
  • Sub-Nyquist ($f_s < 2f_{\max}$): Aliasing (spectral overlap $\rightarrow$ high frequencies fold into lower audible band as distortion).
  • Practical $f_s > 2f_{\max}$: Real Anti-Aliasing Low-Pass Filter (LPF) has finite roll-off / transition guard band (not ideal brick-wall).
  • Bit Depth ($b$): $2^b$ levels $\rightarrow$ smaller step size $\Delta \rightarrow$ lower quantization rounding error / noise floor in quiet passages $\rightarrow$ higher SQNR ($\approx 6\text{ dB/bit}$).

Q3. Communication-Based Watermarking Models OR Robust Watermarking (Unit 5 — 5m)

• Prediction: Compare the 3 Communication-based Watermarking Models with block diagrams, OR explain 4 Robust Watermarking approaches (why Mid-Frequency DCT coefficients are chosen).
• 🔑 Keywords:
  • Basic Model: Cover $c_o =$ channel noise/interference; $c_w = c_o + w_a$; prioritizes robustness; blind detection.
  • Side Information Model: Encoder knows $c_o$ $\rightarrow$ shapes $w_a$ perceptually (stronger in dark/textured regions, weaker in bright/flat); blind detector.
  • Multiplexed Communication Model: $c_o$ (human message) + $m$ (machine message) co-transmitted; balances fidelity & robustness.
  • Robustness 4 Pillars: Redundant Embedding (survives cropping), Spread Spectrum, Perceptually Significant Mid-Frequency Coefficients (HF lost in JPEG quantization; LF degrades visual fidelity), Known Robust Domain (Fourier Magnitude = translation/rotation; DCT = blurring/compression).

Q4. Video GOP Pattern Design OR JPEG Quantization (Unit 2 / Unit 1 — 5m)

• Prediction: Design a GOP pattern for recorded lectures (fast seeking) vs. CCTV surveillance (high compression), OR explain JPEG's 5 stages and why Quantization controls file size.
• 🔑 Keywords:
  • Frames: I-frame (Intra/spatial only, independent anchor, random access/seeking), P-frame (forward motion vectors), B-frame (bi-directional motion estimation, max temporal compression).
  • Lecture VOD: Short GOP (e.g., GOP = 12 or 15, IBBPBBPBBPBB) $\rightarrow$ frequent I-frames $\rightarrow$ fast seeking/scrubbing & error resilience.
  • Surveillance CCTV: Long GOP (e.g., GOP = 48) $\rightarrow$ exploits static background temporal redundancy $\rightarrow$ smaller file size.
  • JPEG: RGB $\rightarrow$ YCbCr $\rightarrow$ Chroma subsampling (4:2:0) $\rightarrow$ Level shift ($-128$) & $8\times8$ DCT $\rightarrow$ Quantization table division (lossy, controls quality vs. bitrate) $\rightarrow$ Zig-zag + RLE + Huffman.

PART B (5 × 10 = 50 Marks)

Q5. Steganography System Design: OutGuess vs. Steghide vs. Model-Based (Unit 5 — 10m)

• Prediction: Design a covert image steganography pipeline for high-surveillance journalists (Cover choice, Selection rule, Embedding principle) and compare OutGuess, Steghide, and Model-Based Steganography.
• 🔑 Keywords:
  • Cachin's Security: $P_C$ (cover distribution) unchanged after embedding ($\text{KL-divergence} = 0$).
  • Cover Work: High-entropy, noisy/textured natural images (landscapes).
  • Selection Rule: Adaptive + Pseudo-random (stego-key seeded) over Sequential.
  • OutGuess (2-Pass DCT): Pass 1 embeds in redundant DCT LSBs; Pass 2 alters remaining unmodified coefficients to restore global DCT histogram (fails if over-embedded).
  • Steghide: Graph-theoretic pair-swapping of values $\rightarrow$ preserves 1st-order histogram natively in 1 pass.
  • Model-Based: Splits cover into Invariant ($C_{\text{inv}}$) + Embeddable ($C_{\text{emb}}$); fits conditional probability $P(C_{\text{emb}} \mid C_{\text{inv}})$.

Q6. Image Forgery Detection Trio + PRNU Pipeline (Unit 3 — 10m)

• Prediction: Given suspect legal evidence photos with Copy-Move, Splicing, and Inpainting (or camera source claim), match and justify the forensic detection pipeline for each.
• 🔑 Keywords:
  • Copy-Move (same image): Keypoint Matching (SIFT / SURF) + Block-based Clone Detection. Invariant to rotation/scaling/illumination (SIFT = accuracy; SURF = speed).
  • Splicing (2 images): PRNU Sensor Noise Correlation ($\text{Corr}(W, \hat{K})$ drops in spliced region) + ELA (resave at known JPEG % and subtract $\rightarrow$ mismatched compression error levels) + Shadow/Sun Azimuth Analysis.
  • Inpainting (object removal): Statistical Analysis (RGB histograms, 1st/2nd-order mean/variance/correlation, local noise/texture discontinuity) + ELA.
  • PRNU 7 Steps: Grayscale $\rightarrow$ Reference Avg ($I_k$) $\rightarrow$ Subtract for Noise Residual ($W_k \rightarrow \hat{K}$) $\rightarrow$ Spatial Filtering $\rightarrow$ Normalization $\rightarrow$ Estimation ($W$) $\rightarrow$ Correlation ($\hat{K}$ vs. $W$).

Q7. Media Server Architecture & Resource Management (Unit 4 — 10m)

• Prediction: Campus media server handling simultaneous live classes + background uploads/maintenance. Analyze CPU, RAM, Disk, and Network management + Pull vs. Push models.
• 🔑 Keywords:
  • Resource Types: Active (CPU, GPU, NIC) vs. Passive (RAM buffer, Disk); Shared vs. Exclusive.
  • Phase 1 — Establishment Phase: Admission Control + Resource Reservation (Guaranteed/Predictable class for live streams; Best-Effort for background uploads).
  • Phase 2 — Runtime Phase: Real-time CPU/Disk scheduling (EDF / Rate Monotonic), policing, Scaling & Adaptation (graceful bitrate/frame-rate downgrade under congestion).
  • 4 Metrics: Throughput, Delay (Local vs. End-to-End), Jitter (smoothed via RAM ring buffers), Reliability (FEC/error concealment).
  • MM Server vs. File Server: Sequential access, real-time deadlines, high concurrency, heavy caching, Push Model (server-driven live class) vs. Pull Model (client-driven VOD lectures).

Q8. Forensic EXIF & Metadata Investigation Case Study (Unit 3 — 10m)

• Prediction: Given an EXIF metadata printout of a suspect photo, identify 5 forensic indicators proving tampering or false location/time/device claims.
• 🔑 Keywords:
  1. GPS Coordinates Mismatch: Lat/Long does not match claimed city (e.g., 34.05 N, 118.24 W = Los Angeles, not San Francisco).
  2. Editing Software Tag: Software: Lightroom Mobile / Photoshop proves post-processing (not raw camera original).
  3. Timestamp Discrepancy: Date Modified is later than Date Created / Date/Time Original.
  4. Timezone / UTC Drift: Compare GPS Timestamp (UTC/Z) against local Date Created to verify local timezone offset at the claimed GPS location.
  5. Device / Optical / Shadow Verification: Cross-check Resolution, Focal Length, F-Stop, Lens Info against claimed smartphone specs + verify sun shadow angles via SunEarthTools.com using EXIF GPS + Timestamp.

Q9. Wardens & Blind vs. Targeted Steganalysis (Unit 5 — 10m)

• Prediction: Cybersecurity agency intercepting social-media images with unknown stego algorithms: (a) Passive, Active, Malicious Wardens; (b) Blind vs. Targeted Steganalysis & Classifier Training.
• 🔑 Keywords:
  • Passive Warden: Inspects silently $\rightarrow$ passes clean traffic or blocks/drops suspected stego (no modification).
  • Active Warden: Modifies all images (re-compression, resizing, slight blurring/noise, stripping metadata) $\rightarrow$ breaks fragile stego payloads without needing to detect them.
  • Malicious Warden: Knows algorithm/keys $\rightarrow$ extracts secret message, alters payload, or impersonates Alice/Bob.
  • Targeted Steganalysis: Algorithm known (e.g., LSB caught via Sum of Absolute Differences [SAD] of adjacent pixels or Chi-Square $\chi^2$ test).
  • Blind / Universal Steganalysis (Winner here!): Algorithm unknown. Extracts cross-domain features (noise residuals, pixel correlations, DCT/Wavelet stats). Trained via Multi-Algorithm Training or One-Class Learning (trained only on clean cover images to flag any statistical outlier).

PART C (2 × 15 = 30 Marks)

Q10. Sliding Window Compression Numerical: LZSS / LZ77 + Decoding / Huffman (Unit 1 — 15m)

• Prediction: String of length $\sim 18\text{–}25$ chars (e.g., MISSISSIPPIMISSISSI or ABRACADABRA...), Window $= 18$, Search $= 12$ (Look-ahead $= 6$).
• 🔑 Keywords / Steps to Show:
  • (a) Encoding Table: 3 columns: Search Buffer (max 12) | Look-Ahead Buffer (max 6) | Output Token.
    • If LZSS: Literal c ($1\text{ B}$) or Pointer (Offset, Length) ($2\text{ B}$).
    • If LZ77: Triplet (Offset, Length, Next_Char).
  • (b) Decoding / Huffman: Step-by-step left-to-right buffer expansion (copy Length chars starting Offset steps back) OR Bottom-up Huffman Tree ($0$ left, $1$ right) + bit count $\sum f_i l_i$.
  • (c) Efficiency: Count Literals ($1\text{ B}$) + Pointers ($2\text{ B}$) $\rightarrow$ Total Bytes $\rightarrow$ Both Compression Ratios ($\frac{\text{Compressed}}{\text{Original}}$ and $\frac{\text{Original}}{\text{Compressed}}$) + $\%$ Space Saved.

Q11. PNG 5-Filter Numerical on an $8 \times 4$ Pixel Matrix (Unit 1 — 15m)

• Prediction: Apply all 5 PNG filters independently to an $8 \times 4$ grayscale matrix and generate 5 separate $8 \times 4$ tables + state which filter is best.
• 🔑 Keywords / Steps to Show:
  • State boundary rule upfront: Out-of-bounds neighbors ($L$ in Col 1, $U$ in Row 1, $UL$ in Row 1 or Col 1) $= 0$, and all negative residuals wrap via $\bmod\ 256$ ($-k \rightarrow 256 - k$).
  • Table 1 (None): Copy input matrix.
  • Table 2 (Sub): $(X - L) \bmod 256$ (Col 1 unchanged).
  • Table 3 (Up): $(X - U) \bmod 256$ (Row 1 unchanged).
  • Table 4 (Average): $(X - \lfloor(L + U)/2\rfloor) \bmod 256$.
  • Table 5 (Paeth): Show 2–3 sample cell workings of $V = L + U - UL$, $|V-L|, |V-U|, |V-UL|$, pick closest neighbor, $(X - P) \bmod 256$.
  • 1-Line Conclusion: The filter producing the smallest, most repetitive residual values maximizes spatial redundancy for the subsequent LZSS + Huffman (DEFLATE) stage.