Slide decks
One deck per paper in the Transformer + FHE collection, written for a student who knows a little machine learning and no cryptography. Every deck follows the same fifteen-slide shape — the problem in plain words, the one big idea, a worked example with real numbers, the threat model, the results, and what the paper does not solve.
Start here — Why transformers are hard to encrypt
CKKS, slots, noise, depth, bootstrapping, and the four hard operators. Every other deck assumes this one.
Hands-on — FHE by hand
Build a homomorphic cipher, break it, then run a small neural network on encrypted inputs. Every box on the slides is live.
Then: the map
The survey gives you the taxonomy; the SoK gives you the honest scorecard of what works today.
Reading the colours
blue = encrypted ·
grey = plaintext ·
orange = the expensive step ·
green = this paper's contribution ·
red = what leaks
Approach
Status
21 of 62 decks written so far. The rest are listed below and are being added module by module.
Start here
2 decks
Why transformers are hard to encrypt
The one deck to read first: CKKS, slots, noise, depth, bootstrapping, and the four operators that make a transformer hard to run under encryption.
15 slides →
FHE by hand — build the cipher, then run a network in it
Hands-on. Invent a homomorphic cipher, break it, fix it, then run a small neural network on encrypted inputs — dragging the sliders that make it fail.
interactive
21 slides →
Surveys & where the field stands
6 decks
A Survey on Private Transformer Inference
The clearest taxonomy of the field — who approximates, who splits the work, and why.
16 slides →
SoK: Private LLM Inference using Approximate Homomorphic Encryption
The honest scorecard — what encrypted LLM inference can actually do today, and at what cost.
16 slides →
Towards Efficient Privacy-Preserving Machine Learning — A Systematic Review
Reviews PPML from three angles at once — protocol, model and system.
16 slides →
A Pragmatic Comparison of Cryptographic Computation Technologies for ML
FHE, MPC and trusted hardware measured against each other on the same tasks.
16 slides →
Comparison of FHE and Garbled Circuit Techniques in PPML Inference
Two very different ways to hide a computation, put side by side.
16 slides →
On the (In-)Security of the Shuffling Defense in Transformer Secure Inference
A popular cheap "defence" is shown not to defend. Read it before trusting a shortcut.
16 slides →
FHE-native, non-interactive inference
12 decks
THE-X — Privacy-Preserving Transformer Inference with HE
The first attempt to push a whole BERT through homomorphic encryption.
FHE-only
18 slides →
Converting Transformers to Polynomial Form for Secure Inference
Replace every non-polynomial piece of a transformer, then train the network to like it.
FHE-only
20 slides →
PowerSoftmax — Towards Secure LLM Inference over Encrypted Data
A softmax replacement designed to be cheap under encryption rather than fitted after the fact.
model redesign
20 slides →
NEXUS — Secure Transformer Inference Made Non-interactive
The landmark non-interactive system — one message each way, client can go offline.
FHE-only
21 slides →
THOR — Secure Transformer Inference with Homomorphic Encryption
Careful matrix packing to cut the rotation bill of encrypted attention.
FHE-only
20 slides →
ELLMo — Packing- and Depth-Aware Encrypted Transformer Inference
Treats packing layout and depth budget as one joint problem instead of two.
FHE-only
19 slides →
Euston — Efficient and User-Friendly Secure Transformer Inference
Non-interactive inference with the client's setup burden deliberately kept small.
FHE-only
18 slides →
STIP — Non-Interactive Transformer Inference via Compact Packing
Compact packing that fits more of the model into each ciphertext.
FHE-only
19 slides →
ATLAS — Automated Approximation of Transformers
Stops hand-tuning polynomial approximations and searches for them automatically.
FHE-only
19 slides →
Private LLM Inference with Homomorphic Encryption (tutorial)
A tutorial walk-through of the whole pipeline — a gentle second stop after the primer.
18 slides →
Fully Homomorphic Encryption on Llama 3
What happens when you point encrypted inference at a modern open LLM.
FHE-only
17 slides →
Secure Transformer Inference Protocol
A protocol-level view of who holds what — and what you get for a permutation instead of a cipher.
18 slides →
Hybrid HE + MPC
16 decks
Iron — Private Inference on Transformers
The paper that opened two-party private transformer inference.
HE + MPC
deck not written yet
BOLT — Privacy-Preserving, Accurate and Efficient Inference
The hybrid baseline everyone else compares against — and the source of the famous round count.
HE + MPC
deck not written yet
BumbleBee — Secure Two-party Inference for Large Transformers
Squeezes the communication cost of the hybrid approach hard.
HE + MPC
deck not written yet
Nimbus — Secure and Efficient Two-Party Inference for Transformers
Moves work into an offline phase so the online phase is short.
HE + MPC
deck not written yet
CipherGPT — Secure Two-Party GPT Inference
The first serious attempt at generative, token-by-token private inference.
HE + MPC
deck not written yet
CipherFormer — Private Inference with Low Round Complexity
Attacks the round count directly — the thing that makes hybrids painful over a real network.
HE + MPC
deck not written yet
East — Efficient and Accurate Secure Transformer Framework
Protocol-level care for the non-linear functions, without giving up accuracy.
HE + MPC
deck not written yet
Primer — Fast Private Transformer Inference on Encrypted Data
Combines several cryptographic tools rather than betting on one.
HE + MPC
deck not written yet
BLB — Breaking the Layer Barrier
Stops treating the transformer layer as an atom and re-cuts the HE/MPC boundary inside it.
HE + MPC
deck not written yet
EncFormer — Secure and Efficient Transformer Inference
Chooses where to put each operation by what packing the next one needs.
HE + MPC
deck not written yet
ENSI — Efficient Non-Interactive Secure Inference for LLMs
Pushes a hybrid-style design towards a single round.
FHE-only
deck not written yet
CipherPrune — Efficient and Scalable Private Transformer Inference
Drops tokens that do not matter — the cheapest ciphertext is the one you never compute.
HE + MPC
deck not written yet
Comet — Private Inference by Predicting Activation Sparsity
Predicts which activations will be zero and skips them. Fast — and a leak worth studying.
HE + MPC
deck not written yet
Reducing Key-Switching Overhead in Two-Party Transformer Inference
A close look at one specific expensive primitive that shows up everywhere.
HE + MPC
deck not written yet
LLMs Can Understand Encrypted Prompt
Reshapes the model to be privacy-computing friendly rather than only the protocol.
HE + MPC
deck not written yet
Powerformer — High-Accuracy Privacy-Preserving Language Modeling
Chases accuracy rather than speed — what does encryption cost in quality?
FHE-only
deck not written yet
Pure MPC baselines
5 decks
MPCFormer — Fast, Performant and Private Transformer Inference
Distils the model into an MPC-friendly one. No FHE at all — useful contrast.
MPC
deck not written yet
PUMA — Secure Inference of LLaMA-7B in Five Minutes
Three-party MPC at LLaMA scale — the speed FHE is measured against.
MPC
deck not written yet
SecFormer — Fast and Accurate Private Inference via SMPC
Keeps accuracy while replacing the awkward non-linear functions.
MPC
deck not written yet
SHAFT — Secure, Handy, Accurate and Fast Transformer Inference
Careful protocol engineering for softmax and GELU under secret sharing.
MPC
deck not written yet
Sigma — Secure GPT Inference with Function Secret Sharing
A different cryptographic tool entirely — function secret sharing, with GPU support.
MPC
deck not written yet
Training & fine-tuning under encryption
5 decks
HETAL — Privacy-preserving Transfer Learning with HE
Encrypted fine-tuning that actually converges — the ICML paper and its extended version together.
FHE-only
deck not written yet
BlindTuner — Encrypted Fine-tuning of Transformers
Fine-tunes a vision transformer while the training data stays encrypted.
FHE-only
deck not written yet
PrivTuner — Homomorphic Encryption with LoRA
Only the small LoRA adapters are trained, so only a little has to be encrypted.
FHE-only
deck not written yet
Encryption-Friendly LLM Architecture
Changes the architecture itself so the hard operators are never there to approximate.
model redesign
deck not written yet
The Inhibitor — ReLU and Addition-Based Attention under TFHE
Attention rebuilt out of additions and ReLU, for a scheme that hates multiplication.
model redesign
deck not written yet
Systems & hardware acceleration
8 decks
EncryptedLLM — LLM Inference via GPU-Accelerated FHE
The same algorithms, but on a GPU — where most of the recent speed came from.
deck not written yet
Cachemir — FHE Inference of Generative LLMs with KV Cache
Generation needs a KV cache. Keeping that cache encrypted is its own problem.
deck not written yet
CryptoGen — Secure Generation with Encrypted KV-Cache Reuse
Reuses encrypted cache entries across tokens instead of recomputing them.
deck not written yet
AEGIS — Long-Sequence HE Transformer Inference on Multi-GPU
Long inputs break the memory budget; this splits the work across GPUs.
deck not written yet
FAME — FPGA Acceleration of Secure Matrix Multiplication
Custom hardware for the one operation encrypted transformers spend most of their time in.
deck not written yet
Chameleon — FHE Scheme Switching Acceleration on GPUs
Different schemes are good at different things — switching between them costs, and this cuts the cost.
deck not written yet
A Scalable Multi-GPU Framework for Encrypted Large-Model Inference
What it takes to scale encrypted inference past one accelerator.
deck not written yet
Dataflow-Oriented Classification of GPU-Accelerated HE
Explains why some GPU HE libraries are fast and others are not.
deck not written yet
Medical & federated
7 decks
MedBlindTuner — Fine-tuning on Biomedical Images with FHE
Encrypted fine-tuning on medical images — the closest paper to what PREMAL is for.
FHE-only
deck not written yet
Privacy-Preserving Federated Vision Transformer in Medical AI
Hospitals train together without sharing patient data, using lightweight HE.
deck not written yet
FedShield-LLM — Secure Federated Fine-Tuning of an LLM
Federated fine-tuning where the updates themselves stay encrypted.
deck not written yet
DictPFL — Private Federated Learning on Encrypted Gradients
Encrypting every gradient is too slow — this encrypts a compact dictionary instead.
deck not written yet
Secure Aggregation using Multiparty Homomorphic Encryption
Nobody holds the whole secret key — the key idea behind federated HE.
deck not written yet
Towards Privacy-Preserving Federated Learning using Hybrid HE
Symmetric encryption on the client, homomorphic work on the server.
deck not written yet
Disposable-key-based Image Encryption for Vision Transformer
Not FHE at all — a cheaper, weaker alternative worth understanding to see what FHE buys.
deck not written yet
Beyond transformers
1 deck
No deck matches that filter.