During a recent project I came accross a common problem when handling content in a frontend. The situation is pretty common, content is stored sowhere on S3 in some structure and users should be able to access it quickly, but only for authorized individuals. Meaning there should be some sort of authentication using some Identity Provider (IDP). In adition the access control should be able to support grouping to bundle access rights for the content.

TL;DR Just give me code


The articles approach uses cookies to store tokens, which are validated using a lambda@edge. Maybe that’s the better approach to take. For my usecase it’s ok to send an access token in the header, for every request. The only relevant thing was: Requests should be cached by Cloudfront as long as the validation succeeds.

Basic solution

As shown in the overview diagram the solution is constructed, using the following service components:

  • Store assets in a secure S3 bucket (default settings, no public access)
  • Fast access to assets on the S3 origin using Cloudfront Distributions as CDN
  • Authentication via OAuth with lambda@edge

On S3 there’s not alot to talk about, just create a bucket with default settings, which blocks all public access. This is what we want, check. Assets should be accessed via Cloudfront. This enables low latency and high transfer speeds. For Cloudfront pricing take a look at this overview.

S3 & Cloudfront

In order to link S3 to your Cloudfront CDN, just create a distribution that resolves the S3 bucket as the distribution’s origin. For now there’s no lambda function to secure our CDN, meaning when you access <dist-url>/path/to/your/ressource-on-bucket.png, you get a response. Ok, now the S3 bucket is effectivly public. Now we need an easy way to secure content on the bucket.


Now we can’t verify the authorization of user. We have no identity provider, to create identities for us. There are many IDPs e.g.: Github, Google, …. Most providers use OAuth2 as authentication flow. An authentifaction flow is basically the process in which an individual proofs the identity to the provider, which involves some process of verification. In case you want to go all the way on this topic see the spec rfc6750.

So Cognito is our identity provider which creates identities for our users, which we can verify by decoding tokens with encryption information (public keys) we obtain from cognito. Like this we can identify an individual as a member of our user pool or group and thus grant or deny access to resources.


Since our distribution exposes the bucket now, we need a security layer on top of this. Here is where lambda@edge comes in. The name tells what it does, executing a lambda function close at the edge of network. There are a few limitations for an edge function:

  • Lambda has to be in us-east-1 (US East, N. Virginia)
  • You can’t configure your Lambda function to access resources inside your VPC
  • Edge function and distribution must be owned by the same AWS account
  • Function code and dependencies (zip) must be < 10MB

The first step is always to find a sample of the event the function has to handle. In this case the following can be used:

    "Records": [
            "cf": {
                "request": {
                    "headers": {
                        "host": [{"value": "d123.cf.net","key": "Host"}],
                        "authorization": [{"value": "Bearer <token>","key": "Authorization"}]
                    "clientIp": "2001:cdba::3257:9652",
                    "uri": "path/to/resource.jpg",
                    "method": "GET"
                "config": {"distributionId": "<dist-id>"}

So the first thing we’re interested in is the request header. There’s our authentication head. For getting started with how such an HTTP header could look like see Mozillas Web/HTTP/Authentication. In our example we want to work with cognito as IDP, since we’re on AWS anyways. Since we’ll obtain some tokens (access, id & refresh token (OAuth2)) from cognito we’ll need the Bearer authentication scheme. At this point we already passed the authentication flow with username/password in the frontend and obtained an access token from cognito, which is sent to cloudfront in the header. The example show that too. You just need to exchange Bearer <token> with the actual token.

Check the header

So let’s check if the header is structured the way we expect:

def check_headers(headers: Dict) -> Union[str, bool]:
    """Check header content for 'Authorization: Bearer <token>' in request
    authorization (key, values).
    if "authorization" not in headers.keys():
        return False
    # Needs only one auth header. Change this when required.
    if not len(headers["authorization"]) == 1:
        return False
    token_string = ""
    auth_values = headers["authorization"][0]["value"].split()
    if not (
        len(auth_values) == 2 and auth_values[0].lower() == "bearer" and auth_values[1]
        return False
    token_string = auth_values[1]
    return token_string

In the example you can see we expect the header to include an authorization part where we also need a bearer followed by the token, split by a blank.

Verify the token

If you want to know more about the JWT token you obtained by cognito, go to jwt.io and test their debugger. Don’t worry the token will never leave jwt.io’s client side. So decoding is done completely in the frontend. I’m not as much into tooling on that topic, but jwt.io is the only place I would paste a token. Anyways when you use their debugger, you’ll see that there are different parts this token consist of.

  • Token header
  • Token payload (data)
  • Tokens verify signature

All the token parts are split by . in the token string and contain different information. The header contains kid (key id) and the encryption algorithm e.g. RS256. To verify if a token comes from our identity provider we take information the IDP gives us and decode the token. So you need the following parameters of your user pool:

  • Region
  • User pool id
  • Client id

Imho the actual decoding should be left to the pros who know what they do - and everybody is lazy anyways. So python-jose is the way to go for decoding JWT tokens. For more detail see the python-jose github-repo. I came across python-jose trying to implement verification with pycognito. A nice project handling the cognito oauth flow for you. Downside was that it’s already too large for the usecase in lambda@edge, since the bundled package doe not pass lambda@edge’s 10 MB restriction. So I went forward taking just the parts required for decoding the access_token. If you want to see where the TokenVerifier code came from have a look at pycognito.

class TokenVerificationException(Exception):
    """Raised when token verification fails, taken from https://github.com/pvizeli/pycognito."""

class TokenVerifier:
    """Simpler version of https://github.com/pvizeli/pycognito."""

    def __init__(
        self, user_pool_id: str, client_id: str, pool_region: str, pool_jwk=None
        self.user_pool_id = user_pool_id
        self.client_id = client_id
        self.pool_region = pool_region
        self.pool_jwk = pool_jwk

    def user_pool_url(self):
        """Construct the user pools jwks url (details at: https://github.com/pvizeli/pycognito)"""
        return f"https://cognito-idp.{self.pool_region}.amazonaws.com/{self.user_pool_id}"

    def user_pool_jwks_url(self) -> str:
        """Construct the user pools jwks url (details at: https://github.com/pvizeli/pycognito)"""
        return f"{self.user_pool_url}/.well-known/jwks.json"

    def get_keys(self) -> Dict:
        """Get public keys from cognitos jwks (details at: https://github.com/pvizeli/pycognito)"""
        if self.pool_jwk:
            return self.pool_jwk
        # If it is not there use the requests library to get it
            self.pool_jwk = requests.get(self.user_pool_jwks_url).json()
        return self.pool_jwk

    def get_key(self, kid) -> str:
        """Get key from pools jwk json 'kids' (details at: https://github.com/pvizeli/pycognito)"""
        keys = self.get_keys().get("keys")
        key = list(filter(lambda x: x.get("kid") == kid, keys))
        return key[0]

    def verify_token(
        self, token: str, id_name: str = "access_token", token_use: str = "access"
    ) -> Dict:
        """Verify token using jwt (details at: https://github.com/pvizeli/pycognito)"""
        kid = jwt.get_unverified_header(token).get("kid")
        hmac_key = self.get_key(kid)
            verified = jwt.decode(
                    "require_aud": token_use != "access",
                    "require_iss": True,
                    "require_exp": True,
        except JWTError:
            raise TokenVerificationException(
                f"Your {id_name!r} token could not be verified."
            ) from None

        token_use_verified = verified.get("token_use") == token_use
        if not token_use_verified:
            raise TokenVerificationException(
                f"Your {id_name!r} token use ({token_use!r}) could not be verified."

        return verified

The decoded token returned by verify_token(...) will be a dictionary, which will hold key-value pairs (decoded token content). At this point we already have everything we want. A header check, token verification and decoding. Everythin we now need to do is further check the tokens content. This could look like this:

# List of user_info in access_token to check
required_user_info = (
    "cognito:groups",  # only required when working with groups

def check_token_access(access_token: Dict) -> bool:
    """Check token contents (user_info and expectation)"""
    # Check if all required user_info is present in token
    if not all(
            (user_info in access_token.keys() and access_token[user_info] is not None)
            for user_info in required_user_info
        return False
    # Check if the token content matches expectation
    if not (
        access_token["token_use"] == "access"
        and access_token["iss"]
        == f"https://cognito-idp.{REGION}.amazonaws.com/{USERPOOLID}"
        and access_token["client_id"] == CLIENTID
        return False
    return True

Further checks

The only thing missing from here on is specific checks, based on the request compaired to the passed token. Here one can do different things here some examples:

  • Match sub (subject) and match it to requested uri
  • Match cognito:group and match it to requested uri
  • Match cognito:<custom-attribute> and match it to requested uri

Some ideas:

def check_authorized_group(
    cloud_front_request: Dict, token: Dict, path_modifier: Callable = hashlib.md5
) -> bool:
    """Checks if 'cognito:group' is path of requested route"""
    token_group = token["cognito:group"]
    if path_modifier:
        token_group = path_modifier(token_group)
    return token_group in cloud_front_request["request"]["uri"]

Wrap it up

Lastly we remove the Authentication header, so Cloudfront can handle the request itself, by passing the input request back to Cloudfront.

def modify_request(cloud_front_request: Dict) -> Dict:
    """Remove header from viewer request for cloudfronts origin request"""
    del cloud_front_request["headers"]["authorization"]
    return cloud_front_request

Here is what the lambda_handler(...) looks like:

REDIRECT_302 = {
    "status": "302",
    "statusDescription": "Found",
    "headers": {"location": HEADERS["location:redirect_uri"]},

# On failure return the following response

def lambda_handler(event, context):
    """Lambda handler"""
    cf_request = event["Records"][0]["cf"]["request"]
    headers = cf_request["headers"]
    header_check = check_headers(headers)
    if isinstance(header_check, bool) and header_check is False:
        print("An error occured (invalid header):")
        return DEFAULT_RESPONE
    token_string = header_check
    token_verifier = TokenVerifier(USERPOOLID, CLIENTID, REGION)
    decoded_token = {}
        decoded_token = token_verifier.verify_token(
            token_string, "access_token", "access"
    except (TokenVerificationException, JWTError, JWSError) as e:
        print("An error occured (token verfication failed):")
        return DEFAULT_RESPONE
    if not check_token_access(decoded_token):
        print("An error occured (unauthorized token content):")
        return DEFAULT_RESPONE
    # Just an example of using the requested path
    if not check_authorized_route(cf_request, decoded_token):
        print("Unauthorized sub in access_token:")
        return DEFAULT_RESPONE
    response = modify_request(cf_request)
    return response

With the REDIRECT_302 response, cloudfront will redirect users to the uri in locations header. This can be the uri pointing to the cognito hosted-ui or frontends login page. From there on users can log in, and the authentication flow will start, until a valid access token is obtained by the user.